Cognitive Neuroscience

The brain in the jar at the Musee Dupuytren

On 17 April 1861 a 51-year-old man died at the Bicetre hospice south of Paris. His name was Louis Victor Leborgne, and for twenty-one years the only syllable he had been able to produce reliably was tan, repeated twice, with a rising or falling intonation that carried most of his meaning. He also swore fluently when provoked. He understood what was said to him, followed instructions, counted on his fingers, and answered questions about how long he had been ill by tapping. The next day Paul Broca opened his skull, found a cavity in the left frontal lobe large enough to hold an egg, and carried the whole brain to the Societe d'Anthropologie de Paris rather than slicing it. That decision is why you can still look at Leborgne's brain in a jar at the Musee Dupuytren, and it is why the most famous lesion in the history of the mind could be re-examined 146 years later with a scanner.

Broca published that year, calling the disorder aphemia and locating the responsible tissue in the posterior third of the left inferior frontal convolution. By 1865, after more cases, he committed to the stronger claim: we speak with the left hemisphere. The region is now called Broca's area, and the syndrome Broca's aphasia, and generations of students have learned a diagram in which a patch of left frontal cortex is labelled speech production.

Key idea: The diagram is a summary of an inference, not an observation. What Broca observed was a dead man who could not speak and a hole in his left frontal lobe. Everything else is reasoning, and the whole of this lesson is about how much weight that reasoning can carry.

What the scanner found in 2007

Nina Dronkers and colleagues put Leborgne's and Lelong's preserved brains in a high-field MRI scanner and reconstructed the lesions in three dimensions without cutting them. The result complicated the textbook. Leborgne's damage did begin in the inferior frontal gyrus, but it did not stop there: it extended into the insula, the surrounding frontal operculum, and deep white matter, including the superior longitudinal fasciculus and the medial subcallosal fasciculus. A surface inspection sees only where the brain has collapsed. The scan sees the tracts that were cut underneath, and Leborgne had lost connections between frontal cortex and the rest of the language network, not merely a patch of cortex.

Two consequences follow. The case that founded localisation is at least partly a disconnection case, an argument for networks rather than seats. And the lesion that gave Broca's area its name does not match what imaging now calls Broca's area: damage confined to pars opercularis and pars triangularis usually yields fluent speech again within months. Persistent Broca's aphasia needs the deeper lesion Leborgne had.

What a lesion actually licenses

The clean statement of lesion logic is narrow. If removing tissue T reliably abolishes ability A, T is necessary for A as the brain is currently organised. That is all. Three claims people slide into are not licensed:

• Sufficiency. Necessity does not mean the tissue performs the function alone. A cut power cable stops a factory without manufacturing anything.
• Localisation of the function. The deficit may arise because T relays, modulates, or connects, not because T computes.
• Specificity. A lesion that abolishes A usually impairs other things too, and the tasks on which the patient is tested are chosen by the examiner.

Against these, five standing problems:

1. Lesions follow blood vessels, not functions. A middle cerebral artery infarct takes out a territory defined by vascular anatomy. Whatever functions happen to share that territory are damaged together, which manufactures correlations between deficits that have nothing to do with shared computation.
2. Diaschisis. Constantin von Monakow's term for the depression of function in intact tissue that has lost its input. Early after a stroke, the deficit overstates the lesion.
3. Reorganisation. Tested at six months, the same patient may be far better, because other tissue has taken over. A deficit measured late understates what the tissue originally did.
4. Selection. Patients who reach a researcher are the ones who survived, were referred, and could cooperate with testing.
5. Task impurity. Every task recruits perception, attention, motivation and response. A patient can fail a memory test because they cannot sustain attention for four minutes.

Karl Lashley spent three decades removing rat cortex in search of the memory trace, found no piece that held it, and proposed equipotentiality. He was wrong, but instructively: his mazes could be solved by vision, touch, smell or kinaesthesis, so removing one route left the rat another. A blunt task plus a blunt lesion yields a blunt answer.

The double dissociation, worked

Suppose patient A fails a memory test and passes a reading test. You would like to conclude that memory and reading use different machinery. You cannot, because there is a dull alternative: the memory test may simply be harder, and A may have a single, general impairment that shows up first on whatever is most demanding. One patient, however many tasks, cannot distinguish a selective deficit from a difficulty gradient.

Now add patient B with the reverse profile: fails reading, passes memory. The difficulty account now has to say that memory is harder than reading for A and reading is harder than memory for B, which is no longer an account of task difficulty at all. This is the double dissociation, named by Hans-Lukas Teuber in 1955, and it is the strongest inference classical neuropsychology offers. Its power is entirely negative: it eliminates the single-resource explanation. It does not tell you what the two systems are.

H.M. and K.F., pointing in opposite directions

On 1 September 1953, William Beecher Scoville removed the medial temporal lobes of a 27-year-old assembly-line worker named Henry Molaison to control intractable epilepsy. The seizures improved. Molaison, known for fifty-five years as H.M., could no longer form new long-term memories of facts or events. Brenda Milner tested him for decades and found the crucial sparing: his digit span was normal at six or seven, and he improved day by day at mirror drawing, a task requiring you to trace a shape while watching your hand in a mirror, while denying each morning that he had ever done it. Immediate span intact, skill learning intact, new declarative memory gone. Lesson 9 develops what that taxonomy became.

Now the other arm. In 1970 Tim Shallice and Elizabeth Warrington described K.F., a young man with a left parieto-occipital lesion whose digit span had collapsed to about two items, yet who learned and retained new verbal material over the long term. H.M. had intact short-term and destroyed long-term storage; K.F. had the reverse. Neither patient alone settles anything. Together they make the single-store model, in which information drips from a short-term buffer into long-term memory, very hard to hold, because K.F.'s buffer was nearly empty and his long-term learning worked anyway.

The point: A double dissociation is built from two patients who are each other's photographic negative. When you read a case report, the first question is always: where is the other half, and has anyone found it?

Where the logic breaks: Plaut's lesioned network

In 1995 David Plaut trained a connectionist network to map the spelling of words onto their meanings. The network had no modules. Concrete and abstract words were handled by the same distributed units, with no boundary anywhere in it. He then damaged it in different places and tested it. Lesions in one part left concrete word reading intact and abstract word reading impaired; lesions elsewhere produced the opposite. A clean double dissociation emerged from a system that contained nothing resembling two separate components.

This is the single most important caveat in classical neuropsychology, and it is not a technicality. The reason the result works is that distributed representations are not uniform: abstract words depend on fewer, more diffuse semantic features, so they degrade differently from concrete words under different patterns of damage. The double dissociation still rules out a single graded resource. It does not license the leap to two boxes in a diagram.

One patient or one hundred

Two research traditions answer the impurity problem in opposite ways. Alfonso Caramazza argued in the 1980s that averaging over patients is incoherent, because a group average over people with different lesions describes no actual cognitive system; the unit of analysis has to be the individual, studied in depth. The group tradition answers that single cases cannot be replicated and invite storytelling, and that with enough patients the accidents of vascular anatomy average out.

The methodological resolution arrived in 2003, when Elizabeth Bates, Stephen Wilson, Nina Dronkers and colleagues published voxel-based lesion-symptom mapping. Instead of sorting patients into syndrome categories, you register every patient's lesion into a common brain space and, at each voxel, compare the behavioural scores of patients whose lesions include that voxel with those whose lesions spare it. The output is a statistical map of where damage predicts a deficit, without anyone having to decide in advance what the syndromes are. It has its own pitfalls, chiefly that vascular architecture biases which voxels can be tested together, but it turned lesion work into something that scales.

Common misconceptions

• Broca's area is where speech is produced. Leborgne's lesion extended through the insula and the underlying white matter, and damage restricted to the cortical region now labelled Broca's area usually produces a transient deficit, not persistent aphasia.
• H.M. lost all his memory. His digit span, vocabulary, childhood memories and motor skill learning were intact. The loss was specific, which is exactly why the case mattered.
• His name was Tan. Tan was the syllable Leborgne could produce. His name appears in the Bicetre records as Leborgne.
• A double dissociation proves two separate modules. Plaut produced one from a network with no modules at all. It rules out a single graded resource and nothing more.
• If a lesion abolishes a function, that function lives there. The tissue may relay or modulate. Necessity is not localisation.

What to carry forward

• Leborgne's brain survives intact because Broca refused to section it, and a 2007 MRI study showed his lesion extended well past the region that now bears Broca's name, into insula and deep white matter.
• A lesion-deficit correlation licenses a claim of necessity as the brain is currently organised, not of sufficiency, localisation, or specificity.
• Vascular anatomy, diaschisis, reorganisation, patient selection and task impurity each distort lesion inference in a predictable direction.
• A single dissociation is consistent with one general impairment plus a difficulty gradient; the second, mirror-image patient is what removes that account.
• H.M. and K.F. form the canonical double dissociation between short-term and long-term memory.
• Plaut's lesioned connectionist network produced a double dissociation without modules, so the inference stops short of licensing boxes in a diagram.
• Voxel-based lesion-symptom mapping replaced syndrome categories with a voxelwise statistical map relating damage to behaviour.

Sources

1. Dronkers, N. F., Plaisant, O., Iba-Zizen, M. T., & Cabanis, E. A. (2007). Paul Broca's historic cases: High resolution MR imaging of the brains of Leborgne and Lelong. Brain, 130(5), 1432-1441. pubmed.ncbi.nlm.nih.gov
2. Scoville, W. B., & Milner, B. (1957). Loss of recent memory after bilateral hippocampal lesions. Journal of Neurology, Neurosurgery and Psychiatry, 20(1), 11-21. pubmed.ncbi.nlm.nih.gov
3. Shallice, T., & Warrington, E. K. (1970). Independent functioning of verbal memory stores: A neuropsychological study. Quarterly Journal of Experimental Psychology, 22(2), 261-273. pubmed.ncbi.nlm.nih.gov
4. Plaut, D. C. (1995). Double dissociation without modularity: Evidence from connectionist neuropsychology. Journal of Clinical and Experimental Neuropsychology, 17(2), 291-321. pubmed.ncbi.nlm.nih.gov
5. Bates, E., Wilson, S. M., Saygin, A. P., Dick, F., Sereno, M. I., Knight, R. T., & Dronkers, N. F. (2003). Voxel-based lesion-symptom mapping. Nature Neuroscience, 6(5), 448-450. pubmed.ncbi.nlm.nih.gov
6. Gazzaniga, M. S., Ivry, R. B., & Mangun, G. R. (2019). Cognitive neuroscience: The biology of the mind (5th ed.), Chapter 3: Methods of cognitive neuroscience. W. W. Norton.

The slide that went in at an angle

In 1958, in a basement laboratory at Johns Hopkins, David Hubel and Torsten Wiesel had an electrode in the visual cortex of an anaesthetised cat and a projector loaded with glass slides carrying black dots. For hours the cell did nothing. Then, pushing a slide into the holder, they heard the loudspeaker attached to the amplifier erupt. The cell was not firing to the dot. It was firing to the faint moving shadow of the slide's own edge as it swept across the screen at a particular angle.

That accident set the agenda for thirty years. The question stopped being whether cortex responds to light and became: what is the specific description of the world that makes this particular cell fire? This lesson is a procedure for answering that question for one neuron, and an honest account of what the answer is worth.

The level below, in one page

A neuron sits at about -65 millivolts at rest. Synaptic input shifts that voltage up (excitatory postsynaptic potentials, mostly glutamate at AMPA receptors) or down (inhibitory, mostly GABA), and the cell sums those shifts across space and time. If the membrane at the axon hillock crosses roughly -55 millivolts, voltage-gated sodium channels open and an action potential fires: about a millisecond long, essentially identical every time, propagating without decrement to the terminals, where it triggers calcium entry and vesicle fusion at a chemical synapse. The biophysics of all this belongs to the companion course BIO 470 Neuroscience, which this course assumes rather than repeats.

Two facts from that level constrain everything above it. The spike is all-or-none, so one spike says nothing about how strong the input was: strength must be coded in rate, in timing, or across a population. And cortical neurons are noisy. Present the same stimulus fifty times and the spike count varies from trial to trial with a variance in the range of the mean, so any claim about what a neuron represents is a claim about a distribution.

Why this matters: The all-or-none spike is the only output a neuron has. Everything cognitive neuroscience says about representation must ultimately be cashed out in which cells fire, how fast, and when.

What the electrode actually hears

In extracellular single-unit recording, a fine-tipped electrode sits in the tissue near a cell body, not inside it. What it measures is the small voltage fluctuation in the extracellular fluid caused by current flowing across nearby membranes: a biphasic waveform lasting about a millisecond, tens to hundreds of microvolts, falling off steeply with distance so that only cells within roughly 50 to 100 micrometres produce isolable spikes. Filter that signal above about 300 hertz and you get spikes; filter below about 300 hertz and you get the local field potential, which reflects summed synaptic activity in a wider neighbourhood.

Because several cells can be in range, the record has to be sorted: waveforms are clustered by shape and amplitude and each cluster treated as one unit. Three sampling biases follow.

• Big cells are over-represented. Large somata and axons generate larger extracellular signals and are found more easily.
• Active cells are over-represented. The experimenter advances the electrode until something fires, so a cell that is silent under the stimuli in use is never entered in the notebook.
• The stimulus set defines the answer. A cell tuned to something nobody thought to present is recorded as unresponsive.

The second point is the problem of dark neurons: whole-cell recording and calcium imaging in awake animals find much lower average rates than the extracellular literature reports, implying that many cortical cells are nearly silent and absent from the classical record.

Mapping one receptive field, step by step

1. Advance and isolate. Lower the electrode in micrometre steps until one waveform stands clearly out of the noise and holds its shape for minutes.
2. Find the field roughly. Sweep a hand-held bar across the screen and listen. The receptive field is the region of visual space in which a stimulus changes the firing.
3. Quantify. Switch to computer-controlled bars at twelve orientations 15 degrees apart, each presented twenty times in randomised order.
4. Build the tuning curve. Plot mean spikes per second against orientation, with error bars.
5. Test the classification. Move the bar within the field and reverse its contrast. Response tied to exact position and polarity means simple; response at the preferred orientation anywhere and to either contrast means complex.
6. Test the negative cases. The wrong orientation and a full-field flash should both fail. A cell that responds to everything has not been characterised.

Here is a tuning curve from step 4, of the kind you would obtain from a typical V1 cell.

Read it. The preferred orientation is near 90 degrees; the half-width at half-height is roughly 25 to 30 degrees, typical for V1; the 4 to 6 spikes per second at the orthogonal orientation is spontaneous rate, not response. What the curve does not tell you: whether the cell is necessary for the cat to discriminate orientation, whether other cells carry the same information, and whether anything downstream reads it.

Simple, complex, and the columns they sit in

Retinal ganglion cells and lateral geniculate cells have concentric centre-surround fields: a small bright spot in the centre excites, the same spot in the surround inhibits, and diffuse light does almost nothing. Simple cells in V1 have elongated fields with separate ON and OFF subregions, which is exactly what you would get by summing a row of aligned centre-surround inputs, and they are therefore selective for orientation and sensitive to the precise position and contrast polarity of the edge. Complex cells keep the orientation preference but discard position and polarity within their field, which is the first step of the invariance that recognition needs.

Hubel and Wiesel then showed how these cells are arranged. Advance an electrode perpendicular to the surface and every cell prefers roughly the same orientation and the same eye. Advance it obliquely and the preference rotates smoothly, about 10 degrees every 50 micrometres, while the dominant eye alternates in bands. A patch about a millimetre square holding a full cycle of orientations for both eyes is a hypercolumn.

Closing one eye for a week

In 1963 Wiesel and Hubel sutured one eyelid of kittens closed for one to three months and then recorded from striate cortex. In a normal cat most V1 cells can be driven through either eye. In the deprived kittens almost none could be driven through the closed eye: the ocular dominance distribution had collapsed onto the open eye, while retina and geniculate stayed largely normal. The same deprivation in an adult cat did almost nothing. Susceptibility was confined to a critical period, peaking in the kitten around the fourth to eighth week.

Cortical organisation is therefore built by competitive activity during development, not specified in advance. The clinical translation is direct: an infant with a dense congenital cataract not operated on within the first months develops permanent amblyopia even after the lens is replaced, and patching a child's stronger eye works less well the later it starts.

How many cells make a percept?

In 1972 Horace Barlow argued for what he called a neuron doctrine for perceptual psychology: that perceptions correspond to the activity of a relatively small number of high-level neurons, and that the properties of single cells are not a level of description beneath psychology but the level at which psychology happens. Pushed to its limit, this becomes the grandmother cell: one neuron that fires for your grandmother and nothing else, so that losing it would delete her.

In 2005 Rodrigo Quiroga, Leila Reddy, Gabriel Kreiman, Christof Koch and Itzhak Fried reported something uncomfortably close. Recording from depth electrodes in the medial temporal lobe of epilepsy patients being monitored for surgery, they found units that responded to one individual across radically different images. One unit in the left posterior hippocampus fired to seven photographs of the actress Jennifer Aniston and not to dozens of other faces, landmarks or animals. Another fired to photographs of Halle Berry, to a line drawing of her, to her in a Catwoman mask, and to her printed name. The response is to the concept, not the pixels.

These are sparse, but they are not grandmother cells, and Quiroga has been consistent about the arithmetic. Each unit responded to a small percentage of the images shown, so no unit is dedicated to one concept; and given how few of the roughly one billion medial temporal lobe neurons any experiment samples, finding several cells for a tested concept implies tens to hundreds of cells per concept.

The core of it: Sparseness is a claim about how many cells respond, not about whether the code is readable from one. Concept cells show that invariant, near-semantic selectivity exists in the human brain. They say nothing about whether any one of them is required for recognition.

What has changed since the 1960s

The one-electrode-one-cell era is over. Neuropixels probes, introduced by James Jun and colleagues in 2017, put 960 recording sites on a shank a few tens of micrometres wide and read out hundreds of neurons at once across several structures. With one cell you ask what it is tuned to; with a thousand you can ask what the population encodes on a single trial, which is a different question and the one most current work is built around.

Common misconceptions

• Neurons are digital, so the brain is a digital computer. The spike is all-or-none, but the variables that carry information (rate, timing, population pattern) are continuous, and dendritic integration below threshold is analogue throughout.
• A complex cell is a simple cell with a larger receptive field. The difference is invariance, not size: a complex cell tolerates changes in position and contrast polarity that a simple cell does not.
• The Jennifer Aniston neuron is a grandmother cell. The same units respond to several concepts, and sampling arithmetic implies many cells per concept.
• If a cell fires to X, the animal is perceiving X. Tuning is a correlation. Establishing that the cell contributes requires interference, which is the subject of lesson 5.
• Recording samples neurons fairly. It over-samples large, active cells, and a cell that is silent under your stimuli is invisible to your method.

The short version

• An extracellular electrode measures spikes from cells within about 50 to 100 micrometres; below 300 hertz the same signal gives the local field potential.
• Recording over-samples large and active neurons, and the stimulus set you choose bounds the tuning you can discover.
• Receptive-field mapping is a procedure: isolate, locate, quantify with randomised presentations, build a tuning curve, then test the classification and the negative cases.
• Simple cells are orientation-selective and sensitive to position and contrast polarity; complex cells keep orientation and discard the rest, which is the first invariance.
• Orientation and ocular dominance are organised in columns, with a hypercolumn analysing one patch of visual field completely.
• Monocular deprivation in kittens shifts ocular dominance drastically during a critical period and hardly at all in adults, which is why congenital cataracts are surgical emergencies in infancy.
• Human medial temporal lobe contains sparse, invariant concept cells; sparse is not the same as single, and tuning is not causation.

Sources

1. Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. Journal of Physiology, 160(1), 106-154. pubmed.ncbi.nlm.nih.gov
2. Wiesel, T. N., & Hubel, D. H. (1963). Single-cell responses in striate cortex of kittens deprived of vision in one eye. Journal of Neurophysiology, 26, 1003-1017. pubmed.ncbi.nlm.nih.gov
3. Barlow, H. B. (1972). Single units and sensation: A neuron doctrine for perceptual psychology? Perception, 1(4), 371-394. pubmed.ncbi.nlm.nih.gov
4. Quian Quiroga, R., Reddy, L., Kreiman, G., Koch, C., & Fried, I. (2005). Invariant visual representation by single neurons in the human brain. Nature, 435(7045), 1102-1107. pubmed.ncbi.nlm.nih.gov
5. Jun, J. J., Steinmetz, N. A., Siegle, J. H., Denman, D. J., Bauza, M., Barbarits, B., et al. (2017). Fully integrated silicon probes for high-density recording of neural activity. Nature, 551(7679), 232-236. pubmed.ncbi.nlm.nih.gov
6. Kandel, E. R., Koester, J. D., Mack, S. H., & Siegelbaum, S. A. (Eds.). (2021). Principles of neural science (6th ed.), Chapters 22 and 24. McGraw Hill.

A psychiatrist with a telepathy hypothesis

In 1924, at the psychiatric clinic in Jena, Hans Berger placed electrodes on the scalp of a young man with a skull defect and recorded a rhythmic oscillation of about ten cycles per second that disappeared when the man opened his eyes. Berger repeated the recording for five years before publishing, in 1929, and his motive was frankly parapsychological: after a near-fatal accident during military service he had come to believe that mental energy could pass between minds, and he was looking for its physical carrier. He found the alpha rhythm instead. Physiologists ignored the paper until Edgar Adrian and Brian Matthews replicated it at Cambridge in 1934, at which point electroencephalography became a method.

Almost a century later the signal Berger found is still the cheapest, fastest window on the living human brain, and its limitations are still the ones he could not have known about. This lesson sets EEG beside its magnetic twin and beside the haemodynamic methods, and asks what each buys.

What is actually being measured

A scalp electrode does not detect spikes. Action potentials are brief, roughly a millisecond, and neighbouring cells fire at unrelated times, so their contributions cancel. What survives to the scalp is the slower postsynaptic current: when glutamate opens channels on the apical dendrite of a cortical pyramidal cell, positive charge flows in there and out near the soma, making the cell a small current dipole that lasts tens of milliseconds.

Three geometric facts follow, and they explain most of what EEG can and cannot do.

• Alignment is everything. Pyramidal cells sit perpendicular to the cortical surface with their apical dendrites parallel to each other, so their dipoles add rather than cancel. Structures with radially symmetric cells, such as the thalamus, contribute almost nothing to the scalp record.
• Synchrony is everything. A single dipole is far too small to detect. Estimates put the requirement at tens of thousands to millions of neurons active within a few milliseconds of one another to produce the 10 to 100 microvolts seen at the scalp.
• The head smears. Current from a cortical patch spreads through cerebrospinal fluid, the highly resistive skull and the scalp. By the time it reaches an electrode it has been blurred across centimetres, which is why adding electrodes improves coverage but does not, by itself, buy sharper localisation.

Remember: EEG is a measure of synchronised synaptic activity in aligned cortical populations. A region can be intensely active and invisible at the scalp if its cells are not aligned or not synchronous.

The same currents, seen magnetically

Every current has a magnetic field around it. In 1968 David Cohen recorded the magnetic counterpart of the alpha rhythm, and magnetoencephalography was born. The fields are tiny, on the order of 50 to 500 femtotesla against a terrestrial background field roughly one hundred million times larger, so MEG needs superconducting quantum interference devices cooled in liquid helium and a magnetically shielded room.

The payoff is that magnetic fields pass through skull and scalp essentially undistorted, so MEG localises better than EEG. The cost is a systematic blind spot. In a spherical head model, a current dipole oriented radially, pointing straight out of the skull, produces no external magnetic field at all. MEG therefore sees sources on the walls of sulci, where cortex is tangential, and is nearly blind to sources on gyral crowns. EEG has no such blind spot. The two methods are complementary rather than redundant, which is why the best source estimates combine them.

The inverse problem, which is not a technical limitation

Given a set of sources inside the head, computing the surface pattern is straightforward physics. Going the other way is not. Hermann von Helmholtz showed in 1853 that the inverse problem has no unique solution: infinitely many internal source configurations produce exactly the same field at the surface. This is a mathematical fact, not a shortcoming of current amplifiers, and no increase in electrode count repairs it.

Practical source estimation therefore adds assumptions and reports results conditional on them: equivalent current dipole fitting assumes a small number of point sources; minimum norm estimation assumes the smallest overall current distribution consistent with the data; beamforming assumes sources are uncorrelated. Each is defensible; none is neutral. When you read a claim that an ERP effect was localised to a region, the honest reading is that the data are consistent with that source under the stated assumptions.

Averaging into components

A single trial of EEG is dominated by activity unrelated to your experiment. The event-related potential is what you get by time-locking to a stimulus and averaging: activity phase-locked to the event survives, everything else shrinks in proportion to the square root of the number of trials. Averaging 100 trials improves the signal-to-noise ratio tenfold, which is why ERP designs need many repetitions and why a small effect can demand several hundred trials per condition.

Components are named by polarity and approximate latency, which is convenient and slightly misleading, since the name refers to a peak while the component is defined by its scalp distribution, its sensitivity and its generator.

The N400 deserves a sentence of its own, because it is the cleanest demonstration in the table. Marta Kutas and Steven Hillyard had people read sentences one word at a time. Most ended sensibly. Some ended as: he spread the warm bread with socks. The anomalous word produced a large negative deflection peaking near 400 milliseconds over centro-parietal sites. The effect is not binary: its amplitude tracks how predictable the word was, so a mildly surprising ending produces a mid-sized N400. That graded sensitivity is what makes it a measure of semantic processing rather than a surprise detector.

Choosing a method

Bottom line: There is no best method, only a match between a question and an instrument. If the question is when, use EEG or MEG. If it is where, use fMRI. If it is whether this tissue is required, neither will answer it, and you need lesson 5.

Rhythms as well as evoked responses

Averaging discards anything not phase-locked to the stimulus, which throws away real signal. Time-frequency analysis keeps it. Alpha activity between 8 and 12 hertz decreases over cortex that is being engaged and increases over cortex that is being suppressed, so alpha lateralisation tracks where spatial attention is directed even before any stimulus arrives. Gamma-band activity above 30 hertz has been linked to feature binding and to attention, though scalp gamma is badly contaminated by tiny eye-muscle artefacts, a problem documented in the late 2000s that invalidated a number of earlier claims.

What EEG is used for when it is not answering a research question

Clinical EEG remains the diagnostic standard for epilepsy and the basis of sleep staging. Mismatch negativity has genuine prognostic value in disorders of consciousness: an unresponsive patient who still generates an MMN is measurably more likely to recover awareness. P300 amplitude to a rare target underpins the classic brain-computer interface speller, in which a user attends to one letter in a flashing grid and the system decodes which row and column evoked the largest P300.

Common misconceptions

• EEG records thoughts, or brain waves that encode mental content. It records the summed synaptic currents of large aligned populations. Decoding specific content from scalp EEG is possible only in constrained tasks and at low information rates.
• More electrodes give better spatial resolution. Denser arrays improve coverage and source modelling, but the blurring by the skull and the non-uniqueness of the inverse problem are not fixed by sampling.
• An ERP component is a single cognitive process. A peak is a sum of overlapping generators. The same P300 label covers a frontal P3a and a parietal P3b with different sensitivities.
• MEG is simply a better EEG. It localises better and is undistorted by the skull, but it is nearly blind to radial sources on gyral crowns, which EEG detects perfectly well.
• A component's latency is the time the brain took to do the process. Latency is the peak of an average; the underlying single-trial latencies vary, and a shift in variability alone changes the averaged peak.

Putting it together

• Berger recorded the human alpha rhythm in 1924 while looking for a physical carrier of telepathy, and the field accepted the result only after Adrian and Matthews replicated it in 1934.
• Scalp EEG reflects postsynaptic currents in aligned, synchronously active pyramidal populations, not action potentials.
• MEG measures the magnetic field of the same currents, passes undistorted through the skull, and is nearly blind to radially oriented sources.
• The inverse problem has no unique solution, so every source localisation is conditional on assumptions such as dipole count, minimum norm or uncorrelated sources.
• Averaging time-locked trials yields ERPs, with noise falling as the square root of trial count; components are labelled by polarity and latency but defined by distribution, sensitivity and generator.
• The N400 is graded by predictability rather than being a binary anomaly detector, which is what makes it a measure of semantic access.
• Match instrument to question: EEG and MEG for timing, fMRI for location, interference methods for necessity.

Sources

1. Cohen, D. (1968). Magnetoencephalography: Evidence of magnetic fields produced by alpha-rhythm currents. Science, 161(3843), 784-786. pubmed.ncbi.nlm.nih.gov
2. Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: Brain potentials reflect semantic incongruity. Science, 207(4427), 203-205. pubmed.ncbi.nlm.nih.gov
3. Sutton, S., Braren, M., Zubin, J., & John, E. R. (1965). Evoked-potential correlates of stimulus uncertainty. Science, 150(3700), 1187-1188. pubmed.ncbi.nlm.nih.gov
4. Berger, H. (1929). Uber das Elektrenkephalogramm des Menschen. Archiv fur Psychiatrie und Nervenkrankheiten, 87, 527-570. (The first published human EEG recordings, including the alpha rhythm and its blocking by eye opening.)
5. Luck, S. J. (2014). An introduction to the event-related potential technique (2nd ed.). MIT Press.

A fish that was definitely not thinking

In 2009 Craig Bennett bought an Atlantic salmon from a fishmonger in Dartmouth, New Hampshire, and put it in a 3 tesla scanner. The salmon was dead. It was shown a series of photographs of people in social situations and asked, in the standard instruction script, to determine what emotion the person in the photograph was experiencing. The data were analysed with a completely ordinary pipeline: motion correction, a general linear model at every voxel, a threshold of p less than 0.001 uncorrected, a minimum cluster size of three voxels. A small cluster of voxels inside the salmon's brain cavity came out statistically significant.

Bennett, Abigail Baird, Michael Miller and George Wolford presented this as a poster at the Human Brain Mapping conference and published it in a joke journal, and it won an Ig Nobel Prize in 2012. It was not a joke about fMRI. It was a demonstration that a pipeline in common use at the time would report activation in tissue that could not possibly be active. This lesson takes that result as a bug report and traces it back through the code.

Step one: what the scanner is measuring

Haemoglobin has a useful magnetic property. Oxygenated haemoglobin is diamagnetic and barely disturbs a magnetic field; deoxygenated haemoglobin is paramagnetic and distorts the field around each red cell, which makes nearby water protons lose phase coherence faster and shortens the decay constant known as T2 star. More deoxyhaemoglobin means a darker image on a T2-star-weighted sequence.

Now the physiology. When a population of neurons becomes more active, local metabolic demand rises, and the vasculature responds by increasing blood flow. The increase in flow overshoots the increase in oxygen consumption, so the local blood becomes more oxygenated than at rest. Deoxyhaemoglobin concentration falls, T2 star lengthens, and the image gets brighter. That is the blood-oxygen-level-dependent signal, discovered by Seiji Ogawa in rodents in 1990 and demonstrated in humans by Ogawa, Kenneth Kwong and Peter Bandettini in 1992.

Three numbers govern how you can use it. The effect is small: 0.5 to 3 percent signal change at 3 tesla for a strong stimulus. It is slow: the response begins about 2 seconds after neural activity, peaks at 4 to 6 seconds, and returns through an undershoot that can last half a minute. And it is spatially coarse relative to neurons: a typical voxel of 3 by 3 by 3 millimetres contains on the order of a million neurons and several hundred thousand synapses per cubic millimetre.

The upshot: fMRI does not measure neural activity. It measures a vascular response that follows neural activity by several seconds, and it measures it as a relative difference between two conditions, because there is no absolute zero for the BOLD signal.

Step two: which neural activity

Nikos Logothetis and colleagues answered this directly in 2001 by recording electrical activity inside the scanner while measuring BOLD from the same monkey visual cortex. The local field potential, which reflects synaptic input and local dendritic processing, predicted the BOLD response better than the spiking output of the same neurons did, and in some conditions the two dissociated.

The implication is uncomfortable and important: BOLD reflects the input to a region and the processing within it more faithfully than what that region sends onward. A region receiving strong inhibitory input can show a robust BOLD increase, because inhibition is itself metabolically expensive. So a bright patch does not mean the region is firing more, and certainly does not mean it is doing the task.

Step three: the arithmetic that killed the salmon

Here is the actual bug. A whole-brain analysis at 3 millimetre resolution runs a separate statistical test at each voxel: on the order of 40,000 to 100,000 tests, one per voxel, each asking whether that voxel's time course matches the task model.

Set the threshold at p less than 0.001 and think about what that means when the null hypothesis is true everywhere. You expect one false positive per thousand tests. With 40,000 voxels, you expect about 40 false positives scattered through the brain, and with 100,000 you expect about 100. In a dead fish, every one of the significant voxels is a false positive by construction, and the salmon's brain cavity contained roughly eight thousand voxels, enough for a few of them to cluster together by chance.

The multiple-comparisons problem has standard fixes, and they differ in what they control.

Bennett and colleagues reported the crucial control: apply either familywise error correction or false discovery rate control to the salmon data, and nothing survives. The dead fish did not break fMRI. It broke uncorrected thresholding, which a survey of the literature at the time found was still being used in a substantial minority of published studies.

Step four: the bug behind the bug

Most researchers had already stopped using uncorrected thresholds. They used cluster-extent inference instead, which is more subtle and was, it turned out, also broken. The logic of cluster inference is that isolated significant voxels are probably noise, while a large contiguous blob is probably real; so you apply a lenient cluster-defining threshold, find contiguous clusters, and ask how likely a cluster that large would be under the null.

Answering that question requires a model of how the noise is spatially correlated. In 2016 Anders Eklund, Thomas Nichols and Hans Knutsson tested the models empirically. They took resting-state scans from 499 healthy people, in whom no task effect can exist, split them into fake two-group experiments, and ran three million group analyses through the standard packages SPM, FSL and AFNI. For a nominal 5 percent familywise error rate, voxelwise inference was conservative, which is safe. Cluster-extent inference with a lenient cluster-defining threshold produced false-positive rates as high as 70 percent. The cause was that the spatial autocorrelation of fMRI noise is not the Gaussian shape the parametric methods assume; it has heavier tails, so large clusters arise by chance far more often than the model predicts. A separate long-standing bug in one package's simulation program made matters worse. Permutation testing, which assumes almost nothing, behaved correctly throughout.

One detail of this paper is itself a lesson in reading. The original significance statement remarked that the results might affect up to 40,000 published fMRI studies, a figure that travelled around the world in headlines. The authors then published a correction in the same journal, because that number was simply the size of the whole fMRI literature, and only a fraction of those papers used the affected form of cluster inference at a lenient threshold. The methodological finding stands. The headline number did not.

So what?: The failure was not fraud and not incompetence. It was an assumption about noise, buried three layers down in a widely trusted piece of software, that nobody had tested against real null data for fifteen years.

Step five: the analytic choices that remain

Two further failure modes survive good thresholding.

Circular analysis. Nikolaus Kriegeskorte and colleagues documented in 2009 that a common workflow selects voxels using one contrast and then reports the effect size in exactly those voxels, from the same data. The selection guarantees an inflated estimate, since you chose the voxels for having a large value. The fix is independent data for selection and testing: a separate localiser run, or cross-validation across scanning runs.

Subtraction and pure insertion. Nearly every fMRI result is a difference between conditions, which assumes that adding a cognitive component to a task leaves everything else unchanged. Donders raised this problem for reaction times in the 1860s and it did not go away when the dependent measure changed. If a harder condition also increases arousal, effort and eye movements, the difference map contains all of that too.

Common misconceptions

• We only use ten percent of our brains. Whole-brain imaging shows activity throughout the brain across a day, and even a small stroke in silent tissue causes deficits. The BOLD maps you see are differences between two conditions with most of the brain subtracted away, which is exactly why they look sparse.
• The region lit up. Nothing lights up. A statistic exceeded a threshold in a comparison between conditions, and the colour scale is chosen by the author.
• fMRI shows where a function happens. BOLD tracks synaptic input and local processing better than output, and cannot distinguish excitation from the metabolic cost of inhibition.
• A significant blob means the effect is large. If the voxels were selected using the same contrast that is then reported, the effect size is inflated by construction.
• Corrected means true. Correction controls false positives under an assumed noise model. Eklund and colleagues showed that when the model is wrong, corrected results can still be wrong at scale.

Where this leaves us

• BOLD is a vascular signal: activity raises blood flow more than oxygen consumption, deoxyhaemoglobin falls, and T2-star-weighted signal rises by roughly 0.5 to 3 percent.
• The haemodynamic response peaks 4 to 6 seconds after the neural event, so fMRI resolves seconds, not milliseconds.
• Logothetis showed that BOLD tracks local field potentials better than spiking, so it reports input and local processing more than output.
• A whole-brain analysis runs tens of thousands of tests; an uncorrected threshold of 0.001 predicts dozens of false positives, which is why a dead salmon produced a significant cluster.
• Bonferroni, random field theory, false discovery rate and permutation control different quantities; permutation makes the fewest assumptions and is now the recommended default.
• Eklund and colleagues showed empirically that parametric cluster-extent inference could reach 70 percent familywise error, and later corrected the widely quoted claim about the number of affected papers.
• Circular voxel selection and the pure insertion assumption remain live problems even when thresholding is done correctly.

Sources

1. Ogawa, S., Tank, D. W., Menon, R., Ellermann, J. M., Kim, S. G., Merkle, H., & Ugurbil, K. (1992). Intrinsic signal changes accompanying sensory stimulation: Functional brain mapping with magnetic resonance imaging. Proceedings of the National Academy of Sciences, 89(13), 5951-5955. pubmed.ncbi.nlm.nih.gov
2. Logothetis, N. K., Pauls, J., Augath, M., Trinath, T., & Oeltermann, A. (2001). Neurophysiological investigation of the basis of the fMRI signal. Nature, 412(6843), 150-157. pubmed.ncbi.nlm.nih.gov
3. Eklund, A., Nichols, T. E., & Knutsson, H. (2016). Cluster failure: Why fMRI inferences for spatial extent have inflated false-positive rates. Proceedings of the National Academy of Sciences, 113(28), 7900-7905. pubmed.ncbi.nlm.nih.gov
4. Kriegeskorte, N., Simmons, W. K., Bellgowan, P. S., & Baker, C. I. (2009). Circular analysis in systems neuroscience: The dangers of double dipping. Nature Neuroscience, 12(5), 535-540. pubmed.ncbi.nlm.nih.gov
5. Bennett, C. M., Baird, A. A., Miller, M. B., & Wolford, G. L. (2010). Neural correlates of interspecies perspective taking in the post-mortem Atlantic salmon: An argument for proper multiple comparisons correction. Journal of Serendipitous and Unexpected Results, 1(1), 1-5.
6. Poldrack, R. A., Mumford, J. A., & Nichols, T. E. (2011). Handbook of functional MRI data analysis. Cambridge University Press.

The problem this lesson exists to solve

Everything in the four preceding lessons is correlational. A tuning curve tells you what a cell prefers, an ERP tells you when a difference appears, a BOLD map tells you where a contrast is significant. None of them tells you whether the brain is using that tissue for the task. A region could be a bystander, receiving copies of information it never acts on. The only way to find out is to change the brain and see whether behaviour changes with it.

Lesions do this, badly and permanently, in whoever happens to have had a stroke. This lesson is about doing it deliberately, reversibly, and on a schedule you control.

A twitching thumb in Sheffield

In 1985 Anthony Barker, Reza Jalinous and Ian Freeston at the University of Sheffield discharged a large capacitor through a coil held against a volunteer's scalp over the motor strip. The volunteer's contralateral hand twitched, painlessly. Electrical stimulation through the scalp had been possible since 1980 but hurt enough to limit its use, because current had to be forced through the skin and skull. Magnetic fields pass through both.

The physics is Faraday induction. Several thousand amperes flow through the coil for about 100 microseconds, producing a magnetic pulse of 1.5 to 2.5 tesla at the scalp. A changing magnetic field induces an electric field in any conductor, and brain tissue is a conductor. The induced field depolarises axons, preferentially where they bend or terminate, and if enough are depolarised together they fire. Over motor cortex that produces a measurable twitch, the motor evoked potential, which gives transcranial magnetic stimulation something rare in human neuroscience: a physiological readout of how much you delivered. Over occipital cortex it produces phosphenes, brief flashes with no light present.

A figure-of-eight coil concentrates the induced field under its centre, giving a focus of roughly a centimetre across at a depth of one to two centimetres. Deeper structures cannot be reached without stimulating everything above them.

The virtual lesion, and what it really is

Vahe Amassian and colleagues ran the experiment that made TMS a cognitive tool. Show a participant a brief array of letters, then fire a single pulse over occipital cortex. If the pulse arrives roughly 80 to 100 milliseconds after the array, the participant cannot report the letters at all. Fire it 40 milliseconds after, or 200 milliseconds after, and report is normal. The letters were seen by the retina in every case. Something in occipital cortex during a specific 50 millisecond window was required for them to reach report.

That is chronometric TMS, and it delivers something no other human method does: a causal claim with a timestamp. Its logic is often summarised as a virtual lesion, and the phrase is worth interrogating.

• TMS does not remove tissue; it adds activity. A pulse drives a burst of largely unstructured firing. The behavioural impairment arises because that noise degrades whatever signal the region was carrying.
• It can therefore help. If a region is currently in a state where added activation pushes a weak representation over threshold, TMS improves performance. Effects depend on the state of the tissue when the pulse arrives.
• It does not stay put. The pulse propagates along the region's connections, so the effective manipulation is on a network.
• It is loud and it is felt. The coil produces a sharp click over 100 decibels and twitches scalp and facial muscles. Any experiment without a sham condition matched for click and sensation is uninterpretable.

What matters here: TMS establishes that a region contributes to a process within a time window. It does not establish what the region computes, and impairment and facilitation are both possible outcomes of the same intervention.

Protocols beyond the single pulse extend the timescale. Repetitive TMS at about 1 hertz tends to depress excitability for tens of minutes; trains at 10 to 20 hertz tend to raise it; theta-burst protocols compress this into 40 seconds. Because trains carry a real seizure risk, published safety guidelines set limits on intensity, frequency and train duration, and screening excludes people with epilepsy or metal implants. Repetitive TMS over left dorsolateral prefrontal cortex is an approved clinical treatment for depression that has not responded to medication.

The cheap alternative, and why to be careful with it

Transcranial direct current stimulation passes 1 to 2 milliamperes of constant current between two sponge electrodes on the scalp. It does not make neurons fire. It shifts resting membrane potential slightly, so that anodal current usually makes firing marginally more likely and cathodal current marginally less. The equipment costs a few hundred pounds, which is why the literature grew very fast and why people build the devices at home.

Two results should govern how you read that literature. In 2015 Jared Horvath, Jason Forte and Olivia Carter meta-analysed single-session tDCS studies in healthy adults and found no reliable effect on any of the many cognitive outcome measures they examined, once they restricted the analysis to effects that had been replicated. In 2018 Mihaly Vöröslakos and colleagues measured intracranial fields directly, in rats and in human cadavers with electrodes in the brain, and found that roughly three quarters of the applied scalp current is shunted through the scalp itself. The field reaching cortex from a typical 1 to 2 milliampere protocol is well under 1 volt per metre, below the level their own recordings showed was needed to affect spike timing measurably.

This does not prove tDCS never does anything. It does mean that a positive result from a small, single-session study, with no measurement of the delivered field, should not move your beliefs much.

Millisecond control of one cell type

The deepest limitation of every method above is that it hits everything in the volume: excitatory and inhibitory cells, axons of passage, glia. Optogenetics removes that limitation, in animals.

Microbial opsins are light-sensitive membrane proteins that move ions. Channelrhodopsin-2, from the green alga Chlamydomonas reinhardtii, opens a cation channel when it absorbs blue light around 470 nanometres. In 2005 Edward Boyden, Feng Zhang, Ernst Bamberg, Georg Nagel and Karl Deisseroth expressed it in mammalian neurons and showed that a flash of blue light produced a single action potential with millisecond reliability, and a train of flashes produced a train of spikes at the frequency they chose. Halorhodopsin, a chloride pump driven by yellow light, does the reverse and silences cells.

Three properties make this a different class of tool. Specificity: the opsin is delivered by a viral vector under a cell-type-specific promoter, or in a mouse line where expression depends on a recombinase driven by a particular gene, so only parvalbumin interneurons, or only dopamine neurons projecting to one target, carry the protein. Timing: control is at the millisecond scale, matching the timescale of the circuit. Reversibility: the manipulation ends when the light goes off.

The first clinical use came in 2021, when Jose-Alain Sahel and colleagues delivered an opsin to retinal ganglion cells in a patient blinded by retinitis pigmentosa and, with light-amplifying goggles, restored the ability to locate and count objects on a table.

The catch nobody expected

In 2015 Timothy Otchy, Steffen Wolff, Bence Olveczky and colleagues compared acute and chronic manipulations of the same structure in songbirds and rodents. Acutely silencing a cortical nucleus disrupted song severely. Permanently lesioning the same nucleus, and allowing recovery, left song essentially intact. The acute manipulation was not revealing that the nucleus was necessary; it was injecting a disruptive perturbation into downstream circuits that were still listening to it.

This is a general problem for every interference method, TMS included. A brain is a recurrent system, and abruptly changing one node's output disturbs everything reading from it. Acute and chronic results dissociating does not mean one of them is a mistake. It means they answer different questions: what happens when this node's signal is corrupted right now, and what the system can do without this node at all.

Common misconceptions

• TMS switches a region off. It adds unstructured activity. Depending on the state of the tissue it can impair or improve performance.
• tDCS enhances cognition. The largest quantitative review of single-session studies in healthy adults found no reliable replicated effect, and direct measurement shows most of the current never reaches cortex.
• Optogenetics is used on human brains. Outside a single retinal therapy, it is an animal method. It requires gene delivery and an implanted light source.
• An acute silencing result is a cleaner lesion. Acute and chronic manipulations of the same structure can give opposite answers, because acute perturbation corrupts what downstream circuits are still reading.
• Phosphenes prove the pulse reached the visual cortex precisely. Phosphenes confirm that excitable visual tissue was activated somewhere in the induced field, which can include the optic radiation.

What you now know

• Correlational methods cannot establish that tissue is used; interference can, which is why this toolkit exists.
• TMS induces current by Faraday induction, focally to about a centimetre and a depth of one to two centimetres, with motor evoked potentials providing a dosimetry readout.
• Chronometric TMS gives causal claims with timestamps: an occipital pulse 80 to 100 milliseconds after a letter array abolishes report of the letters.
• The virtual lesion metaphor is loose: TMS injects noise, propagates through connections, can facilitate, and requires sham control for click and scalp sensation.
• tDCS delivers well under 1 volt per metre to cortex once scalp shunting is measured, and the largest review of single-session cognitive effects in healthy adults found none that replicated.
• Optogenetics couples cell-type specificity, millisecond timing and reversibility, and has one clinical application so far, in the retina.
• Acute perturbation and permanent lesion can dissociate, as Otchy and colleagues showed in songbirds; they answer different questions rather than one being wrong.

Sources

1. Barker, A. T., Jalinous, R., & Freeston, I. L. (1985). Non-invasive magnetic stimulation of human motor cortex. The Lancet, 325(8437), 1106-1107. pubmed.ncbi.nlm.nih.gov
2. Walsh, V., & Cowey, A. (2000). Transcranial magnetic stimulation and cognitive neuroscience. Nature Reviews Neuroscience, 1(1), 73-79. pubmed.ncbi.nlm.nih.gov
3. Horvath, J. C., Forte, J. D., & Carter, O. (2015). Quantitative review finds no evidence of cognitive effects in healthy populations from single-session transcranial direct current stimulation. Brain Stimulation, 8(3), 535-550. pubmed.ncbi.nlm.nih.gov
4. Vöröslakos, M., Takeuchi, Y., Brinyiczki, K., Zombori, T., Oliva, A., Fernandez-Ruiz, A., et al. (2018). Direct effects of transcranial electric stimulation on brain circuits in rats and humans. Nature Communications, 9(1), 483. pubmed.ncbi.nlm.nih.gov
5. Boyden, E. S., Zhang, F., Bamberg, E., Nagel, G., & Deisseroth, K. (2005). Millisecond-timescale, genetically targeted optical control of neural activity. Nature Neuroscience, 8(9), 1263-1268. pubmed.ncbi.nlm.nih.gov
6. Otchy, T. M., Wolff, S. B., Rhee, J. Y., Pehlevan, C., Kawai, R., Kempf, A., et al. (2015). Acute off-target effects of neural circuit manipulations. Nature, 528(7582), 358-363. pubmed.ncbi.nlm.nih.gov

A slot, a card, and a woman who could not say which way it pointed

In 1988 a 34-year-old woman known in the literature as D.F. was overcome by carbon monoxide from a faulty gas heater while showering. She survived. Her acuity, colour vision and visual fields recovered largely intact, and she could tell you the colour and surface texture of anything you handed her. What she could not do was see shape. Shown a simple line drawing, she could not name it or copy it. Shown a slot cut in a disc, and asked to rotate a hand-held card until it matched the slot's orientation, she performed at close to chance, scattering her answers across the full range.

Then David Milner and Melvyn Goodale asked her to post the card through the slot. She reached out and rotated her wrist smoothly as she went, arriving with the card aligned, and posted it as accurately as a control participant. The same orientation information, in the same brain, at the same moment: unavailable to report and completely available to the hand.

Key idea: Vision is not one system delivering one representation that everything downstream shares. D.F. is the demonstration that a visual property can be computed and used for action while being absent from what the person can see.

Getting the retina onto the cortex

Before the streams, the map. Light is transduced in the retina, ganglion cell axons leave through the optic nerve, and at the chiasm the fibres from each nasal hemiretina cross. From there the two optic tracts each carry the contralateral half of the visual field. Most fibres end in the lateral geniculate nucleus, a six-layered structure in the thalamus, where two large ventral layers carry the magnocellular pathway (fast, achromatic, sensitive to low contrast and high temporal frequency) and four dorsal layers carry the parvocellular pathway (slower, colour-opponent, high spatial resolution).

From the geniculate the projection goes to primary visual cortex in the calcarine sulcus, and it goes there in an orderly way. Retinotopy means that neighbouring points on the retina project to neighbouring points in cortex. The map is not uniform. Cortical magnification means that the central few degrees of the visual field are allotted a wildly disproportionate share of cortical area: the central 10 degrees, which is a little over one percent of the visual field by area, occupies more than half of V1. Fovea-heavy vision is not a property of the retina alone but of how much cortex is spent on it.

Michael Sereno, Anders Dale and colleagues made this measurable in humans in 1995 with phase-encoded mapping. A participant watches a wedge rotating slowly around fixation and an annulus expanding outward. Each voxel's BOLD time course peaks at the moment its preferred polar angle or eccentricity is stimulated, so the phase of the response gives the location it represents. Where the polar-angle map reverses direction, you have found a border between visual areas, and this is still how V1, V2, V3, hV4 and MT are defined in an individual brain.

Two streams, first version

In 1982 Leslie Ungerleider and Mortimer Mishkin proposed, on the basis of monkey lesion work, that cortical vision divides into two pathways leaving occipital cortex. A ventral stream running into inferotemporal cortex handles object identity; a dorsal stream running into posterior parietal cortex handles spatial location. Monkeys with inferotemporal lesions failed at telling one object from another but did well on a landmark task requiring them to judge which of two positions was closer to a cue; monkeys with parietal lesions showed the opposite. What and where.

The scheme is elegant and it is not quite right, and D.F. is why.

Two streams, second version

Milner and Goodale's reframing in 1992 kept the anatomy and changed the job description. The distinction, they argued, is not between two kinds of information but between two uses of information. The ventral stream builds a durable, allocentric representation suitable for recognition, comparison and report. The dorsal stream computes egocentric, moment-to-moment parameters for the control of action: how far, how big for my grip, what orientation for my wrist. Vision for perception and vision for action, rather than what and where.

The evidence is a double dissociation of exactly the shape lesson 1 described.

The condition in the right-hand column, optic ataxia, is not a general clumsiness. The patient can reach accurately to a part of their own body with their eyes closed and can describe the target object in detail. What fails is the visual guidance of the reach itself, and characteristically it fails worse in peripheral vision than at fixation.

The point: The pair rules out the dull account, that D.F.'s posting is preserved because posting is easier than reporting. For the optic ataxia patient, reporting is the easy one.

The objection you should hold on to

A widely cited extension of the perception-action idea claimed that grasping is immune to size illusions: presented with the Ebbinghaus display, where a disc looks larger when ringed by small circles, people supposedly report the illusion but scale their grip to the true size. If true, that is a beautiful demonstration that the hand is served by a stream that the illusion cannot reach.

It has not held up cleanly. Volker Franz and colleagues showed that the perceptual and grasping measures in these studies were not comparable: the perceptual task typically asked for a comparison of two discs while the grasp was directed at one, and when the two tasks were matched properly the grip showed much the same illusion effect as perception. The current position among most researchers is that the anatomical dissociation in patients is robust, and the strong claim that action is illusion-proof in healthy people is not.

Two further qualifications. The streams are massively interconnected, not parallel pipes; and D.F.'s spared ability is best for immediately available targets, degrading when a delay of a few seconds is introduced, which suggests the dorsal system does not store.

Motion, and a patient who saw the world in stills

Area MT, also called V5, is a small region on the dorsal side of the occipitotemporal junction in which the great majority of cells are direction-selective, organised in columns by preferred direction. Two results tie it to seeing motion.

The causal one, from William Newsome's laboratory in 1990: a monkey judges the net direction of a noisy field of moving dots while the experimenters pass a tiny current through a column of MT cells with a known direction preference. The monkey's judgements shift toward that preferred direction, by an amount equivalent to adding real motion signal to the display. Stimulating a few hundred cells in the right column biases what the animal reports seeing.

The clinical one, from Josef Zihl and Detlev von Cramon in 1983: patient L.M., after bilateral damage in this region, lost the perception of motion while keeping acuity, colour and object recognition. She described pouring tea as watching the liquid appear frozen, like a glacier, with the cup suddenly full. Crossing a road became dangerous because a car that had been distant was abruptly close. Motion, it turns out, is a visual attribute that can be selectively destroyed.

Common misconceptions

• The eye sends a picture to the brain. The retina sends about a million ganglion cell outputs coding local contrast, colour opponency and temporal change, already heavily processed, and the cortical map devotes over half its area to the central few degrees.
• The dorsal stream is the where stream. Location is computed in both. The better-supported distinction is what the information is for: durable representation for recognition, versus egocentric parameters for an action happening now.
• D.F. is blind. Her acuity, colour vision and visual fields are largely intact. What she lacks is the perception of form, which is why her case is informative rather than merely severe.
• Action is immune to visual illusions. When perceptual and grasping tasks are matched properly, grip aperture shows illusion effects of much the same size as perceptual report.
• Damage to a visual area removes a place, not a property. L.M. lost motion perception specifically, with acuity, colour and recognition intact, which is a loss of an attribute.

Recap

• Retinotopic maps preserve neighbourhood relations but not scale: over half of V1 represents the central 10 degrees.
• Phase-encoded fMRI with rotating wedges and expanding rings defines visual area borders at reversals of the polar-angle map, in individual brains.
• Ungerleider and Mishkin divided cortical vision into what and where on the basis of monkey lesion dissociations.
• Milner and Goodale reframed the division as perception versus action, with the ventral stream building representations for recognition and the dorsal stream computing parameters for immediate movement.
• D.F. and optic ataxia patients form the double dissociation: reporting slot orientation versus posting a card through it, in opposite directions.
• The strong claim that grasping escapes size illusions did not survive properly matched comparisons, though the patient dissociation stands.
• Microstimulation of a direction column in MT biases what a monkey reports seeing, and bilateral damage to the same region produced selective loss of motion perception in patient L.M.

Sources

1. Goodale, M. A., Milner, A. D., Jakobson, L. S., & Carey, D. P. (1991). A neurological dissociation between perceiving objects and grasping them. Nature, 349(6305), 154-156. pubmed.ncbi.nlm.nih.gov
2. Sereno, M. I., Dale, A. M., Reppas, J. B., Kwong, K. K., Belliveau, J. W., Brady, T. J., et al. (1995). Borders of multiple visual areas in humans revealed by functional magnetic resonance imaging. Science, 268(5212), 889-893. pubmed.ncbi.nlm.nih.gov
3. Salzman, C. D., Britten, K. H., & Newsome, W. T. (1990). Cortical microstimulation influences perceptual judgements of motion direction. Nature, 346(6280), 174-177. pubmed.ncbi.nlm.nih.gov
4. Zihl, J., von Cramon, D., & Mai, N. (1983). Selective disturbance of movement vision after bilateral brain damage. Brain, 106(2), 313-340. pubmed.ncbi.nlm.nih.gov
5. Milner, A. D., & Goodale, M. A. (2006). The visual brain in action (2nd ed.). Oxford University Press.
6. Franz, V. H., Gegenfurtner, K. R., Bulthoff, H. H., & Fahle, M. (2000). Grasping visual illusions: No evidence for a dissociation between perception and action. Psychological Science, 11(1), 20-25.

A patient watching a doctor's face melt

In 2012 a man being monitored with electrodes on the surface of his right fusiform gyrus, in preparation for epilepsy surgery, was looking at Josef Parvizi's face when the team passed a small current through two adjacent electrodes. He said, immediately: you just turned into somebody else, your face metamorphosed, your nose got saggy and went to the left. When the same current was applied while he looked at objects rather than a face, nothing happened. The electrodes sat in tissue that had already been shown, by fMRI in the same patient, to respond more to faces than to other categories.

That result is about as clean a causal demonstration as human cognitive neuroscience produces, and it is one exhibit in an argument that has run for a quarter of a century: is there a piece of cortex whose job is faces, or is there a piece of cortex that ends up doing faces because faces are what we are all experts at?

The claim

In 1997 Nancy Kanwisher, Josh McDermott and Marvin Chun scanned people while they viewed faces and everyday objects, and found a patch in the right mid-fusiform gyrus responding roughly twice as strongly to faces. It appeared in most participants, in approximately the same place, and it survived a battery of controls: faces beat scrambled faces, houses, hands, and objects matched for low-level properties. They named it the fusiform face area, and argued it constitutes evidence for a cortical module dedicated to a domain, in the sense that evolution has built machinery for a specific and evolutionarily important class of stimulus.

The supporting case has four legs.

• Behavioural signatures. Faces are recognised holistically. Turn a face upside down and recognition collapses far more than it does for other objects, a disproportion that is hard to obtain with any other category.
• Single cells. Doris Tsao, Winrich Freiwald, Roger Tootell and Margaret Livingstone used fMRI in macaques to find a face-selective patch, then put electrodes into it. Of the visually responsive cells they recorded, the overwhelming majority, around 97 percent, were face-selective. A region defined by a blurry haemodynamic contrast turned out to be almost purely face cells.
• Causal evidence. Parvizi's stimulation result above, which distorted the perception of a face and left objects alone.
• Patients. Damage to this region and its neighbours produces prosopagnosia. In 1993 Jane McNeil and Elizabeth Warrington described a prosopagnosic man who took up sheep farming and became able to identify individual sheep by their faces, while remaining unable to identify people by theirs.

The counter-claim

Isabel Gauthier and Michael Tarr argued that the region is not for faces but for the fine discrimination of individuals within a highly familiar category, and that faces dominate it because everyone has spent a lifetime practising them. In 1999 they trained participants for many hours on greebles, invented three-dimensional objects with a shared configuration and individually varying parts, and showed that the same middle fusiform region increased its response to greebles as expertise developed. They and others reported the same for car experts viewing cars and bird experts viewing birds.

The stronger version of the objection came from James Haxby and colleagues in 2001, and it changed how the whole field analyses data. Instead of asking which voxel responds most to faces, they asked whether the pattern of activity across ventral temporal cortex could identify what a person was looking at. It could, for eight categories. More pointedly, when they removed the voxels that responded maximally to faces, the remaining pattern still classified faces well above chance; and the pattern inside the face-selective region still carried information distinguishing chairs from shoes. Category information is distributed and overlapping, not confined to a module.

In short: Both sides accept the data. They disagree about what makes a region special: what it is built for, or what it has become good at.

The developmental evidence, which cuts across both

In 2017 Michael Arcaro, Peter Schade, Justin Vincent, Carlos Ponce and Margaret Livingstone reared macaques with human handlers who wore welding masks, so the animals grew up seeing hands, bodies and objects but no faces. Scanned as juveniles, these monkeys had normal category-selective patches for hands and for objects, and no face patches at all. That is strong evidence that experience is required to build the domain.

But the same laboratory showed the patches do not land anywhere. Face patches consistently develop in retinotopic territory biased toward central vision, body patches in territory biased toward the periphery, and that proto-organisation is present in infancy before the category selectivity appears. So the honest summary is neither of the two clean positions: a coarse map exists from the start, and experience carves the categories into it.

What prosopagnosia does and does not show

Acquired prosopagnosia follows damage to the right or bilateral occipitotemporal cortex, typically from posterior cerebral artery stroke or herpes encephalitis. Patients see faces perfectly well. They describe eyes, note a moustache, judge age, sex and expression, and simply cannot say whose face it is, sometimes including their own in a mirror. They compensate with voice, gait, hairstyle and context.

Developmental prosopagnosia occurs without any lesion, runs in families, and in a large German sampling by Ingo Kennerknecht and colleagues was estimated at roughly 2 percent of the population, which means several people in any lecture hall. At the other end, super-recognisers identify faces seen once, briefly, years earlier.

The critical question for the dispute is whether the deficit is face-specific or a general failure of within-category individuation. The evidence is genuinely mixed. Some patients, like Warrington's sheep farmer, appear face-specific. Others fail equally on cars, birds and flowers when the task is made properly comparable in difficulty, which is exactly the methodological trap lesson 1 described: the face task and the car task are almost never matched.

Worth holding on to: A face-specific deficit requires showing intact performance on a non-face task of equal difficulty in the same patient. Very few reported cases meet that standard, which is why the literature stays unsettled.

What would settle it

Elizabeth McKone, Kanwisher and Bradley Duchaine set out the terms in 2007, and they are still the right ones.

1. Match the tasks. Expertise effects and face effects must be compared with equal discriminability and equal difficulty, or the comparison is uninterpretable.
2. Test the causal claim on both. If the region is an expertise engine, disrupting it should impair a car expert's car discrimination as much as face identification.
3. Look at infants. Newborn preferences for face-like configurations, and the very early emergence of face-selective responses, put a hard limit on how much expertise can explain.
4. Ask what a module would even predict. A cortical region can be domain-specific in what it is wired to receive and still be sculpted by experience. The two positions are not exclusive, and the sharp version of each is now held by fewer people than the argument's reputation suggests.

Common misconceptions

• The FFA is where faces are recognised. It is one node in a network including the occipital face area, the posterior superior temporal sulcus for dynamic aspects, and anterior temporal regions for identity, and damage to any of them can impair recognition.
• Prosopagnosics see faces as blurs. Perception is intact. They can describe a face in detail and cannot say who it is.
• Everyone's face area is in the same voxels. Location varies enough between people that studies define it individually with a functional localiser, which is also why group averaging blurs it.
• Prosopagnosia is always caused by brain damage. The developmental form has no lesion, runs in families, and affects on the order of one person in fifty.
• A region that responds twice as strongly to faces is a face detector. Response magnitude is a comparison between conditions; pattern analyses show the same region carries information about non-preferred categories too.

The takeaway

• Kanwisher, McDermott and Chun defined a right mid-fusiform region responding about twice as strongly to faces as to objects, consistently across participants.
• Tsao and colleagues showed that a macaque face patch is composed almost entirely of face-selective cells, and Parvizi's stimulation distorted face perception specifically.
• Gauthier and Tarr showed that expertise with greebles, cars or birds raises the response in the same region, supporting an individuation-expertise account.
• Haxby's multivariate analysis showed category information distributed across ventral temporal cortex, including outside the face-selective voxels.
• Monkeys reared without seeing faces develop no face patches, yet patch locations are constrained by a retinotopic proto-organisation present before the selectivity emerges.
• Acquired and developmental prosopagnosia leave perception of the face intact and remove identification; developmental cases affect roughly 2 percent of people.
• The dispute turns on task matching: demonstrating a face-specific deficit requires an equally difficult non-face control task in the same patient.

Sources

1. Kanwisher, N., McDermott, J., & Chun, M. M. (1997). The fusiform face area: A module in human extrastriate cortex specialized for face perception. Journal of Neuroscience, 17(11), 4302-4311. pubmed.ncbi.nlm.nih.gov
2. Gauthier, I., Tarr, M. J., Anderson, A. W., Skudlarski, P., & Gore, J. C. (1999). Activation of the middle fusiform face area increases with expertise in recognizing novel objects. Nature Neuroscience, 2(6), 568-573. pubmed.ncbi.nlm.nih.gov
3. Haxby, J. V., Gobbini, M. I., Furey, M. L., Ishai, A., Schouten, J. L., & Pietrini, P. (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science, 293(5539), 2425-2430. pubmed.ncbi.nlm.nih.gov
4. Tsao, D. Y., Freiwald, W. A., Tootell, R. B., & Livingstone, M. S. (2006). A cortical region consisting entirely of face-selective cells. Science, 311(5761), 670-674. pubmed.ncbi.nlm.nih.gov
5. Parvizi, J., Jacques, C., Foster, B. L., Withoft, N., Rangarajan, V., Weiner, K. S., & Grill-Spector, K. (2012). Electrical stimulation of human fusiform face-selective regions distorts face perception. Journal of Neuroscience, 32(43), 14915-14920. pubmed.ncbi.nlm.nih.gov
6. Arcaro, M. J., Schade, P. F., Vincent, J. L., Ponce, C. R., & Livingstone, M. S. (2017). Seeing faces is necessary for face-domain formation. Nature Neuroscience, 20(10), 1404-1412. pubmed.ncbi.nlm.nih.gov
7. McKone, E., Kanwisher, N., & Duchaine, B. C. (2007). Can generic expertise explain special processing for faces? Trends in Cognitive Sciences, 11(1), 8-15. pubmed.ncbi.nlm.nih.gov

A cathedral square described from two directions

In 1978 Edoardo Bisiach and Claudio Luzzatti asked two Milanese patients with right parietal damage to imagine standing on the steps of the cathedral, looking out over the Piazza del Duomo, and to name the buildings they could see. Each patient listed the buildings on the right side of that imagined view and none on the left. Then the experimenters asked them to imagine standing at the far end of the piazza, facing the cathedral, and to describe it again. The patients now named the buildings they had just omitted, and omitted the ones they had just named.

Nothing was wrong with their eyes; they had their eyes closed. Nothing was wrong with their memory for Milan, since between them the two descriptions covered the square. What failed was the ability to attend to the left of a represented space, whatever space that happened to be. This lesson works through the experimental machinery that makes that result interpretable.

Procedure one: the cueing paradigm, step by step

Michael Posner's task is three decades of attention research compressed into eight seconds per trial.

1. The participant fixates a central cross, with two boxes to left and right.
2. A cue appears: either one box brightens (a peripheral cue) or an arrow at fixation points one way (a central, symbolic cue).
3. After a stimulus onset asynchrony of, say, 100 milliseconds, a target dot appears in one of the boxes.
4. The participant presses a key as soon as they detect the target, without moving their eyes. Eye position is monitored, because the whole point is covert attention.
5. Cues are valid on 80 percent of trials and invalid on 20 percent, with neutral trials as a baseline.

The result is a validity effect of roughly 20 to 50 milliseconds: valid trials are faster than neutral, invalid trials slower. Attention can be moved to a location without moving the eyes, and doing so speeds processing there at a cost elsewhere.

Now change one input and watch the result flip. Keep the peripheral cue and lengthen the interval to 300 milliseconds or more. The advantage reverses: responses to targets at the cued location become slower than at the uncued location. This is inhibition of return, and it is thought to bias attention toward unexplored locations, which is useful when searching. Do the same with the central arrow cue and no reversal occurs; the benefit simply grows for the first few hundred milliseconds and holds.

Why this matters: Two orienting systems with different time courses can be separated with nothing more than a cue type and a delay. This is what a good paradigm buys: a manipulable handle on a process that is otherwise invisible.

Procedure two: search, and what attention is for

Anne Treisman's visual search task varies the number of items in a display and measures how search time grows with that number. Two patterns emerge. Looking for a red bar among green bars, or a vertical bar among horizontal ones, gives a search time that hardly changes as items are added: the target pops out, and the slope of the function is close to flat. Looking for a red vertical bar among red horizontal bars and green vertical bars gives a slope of roughly 20 to 30 milliseconds per item, and about twice that when the target is absent, as though items were being checked one at a time.

Treisman's feature integration theory explained the difference. Simple features such as colour, orientation and motion are registered in parallel across the visual field in separate feature maps. What is not registered in parallel is which feature belongs to which object. Binding a red to a vertical requires attention to be directed to the location where they coincide, so a conjunction search proceeds item by item.

The theory made a strange prediction that proved correct. If attention is the glue, then withdrawing attention should let features migrate. Treisman and Hilary Schmidt showed brief displays with a demanding task at fixation, and participants reported illusory conjunctions: shown a red X and a blue O in the periphery, they confidently reported seeing a blue X. Not a guess, not a failure to see: a confident percept of a combination that was never presented.

The core of it: Features and their bindings are separable. The binding problem is real, and attention is at least part of the answer.

The main objection, from John Duncan and Glyn Humphreys, is that the two search modes are not two mechanisms but two ends of a continuum. Search efficiency varies smoothly with how similar the target is to the distractors and how similar the distractors are to each other. Make a feature search hard enough, by reducing the colour difference, and the slope steepens; make a conjunction search easy enough and it flattens. Feature integration theory survives as an account of binding rather than as a taxonomy of searches.

What attention does to a neuron

Jeffrey Moran and Robert Desimone put the question to single cells in monkey area V4 in 1985. They found a cell's receptive field, placed two stimuli inside it, one that the cell liked and one it did not, and trained the animal to attend to one or the other while holding fixation. The cell's response depended on which stimulus was attended: attend the preferred one and the cell responded strongly, attend the poor one and the response collapsed toward what the poor stimulus alone would produce. The unattended stimulus was largely filtered out, as if the receptive field had shrunk around the attended object.

Desimone and Duncan generalised this in 1995 as biased competition. Objects in a visual scene compete for representation, because neurons at higher levels have large receptive fields containing several objects. Attention is not a separate spotlight added to the system; it is a bias signal, from prefrontal and parietal sources, that tips the competition toward one object. Crucially, this predicts that attention effects should be larger when two objects share a receptive field than when only one is present, which is what is found.

Two networks, and what happens when one breaks

Maurizio Corbetta and Gordon Shulman's 2002 synthesis divides human attention into two frontoparietal systems. A dorsal network, centred on the intraparietal sulcus and frontal eye fields, carries goal-directed selection: it holds the current priority map and biases sensory cortex accordingly. A ventral network, in the right temporoparietal junction and ventral frontal cortex, detects behaviourally relevant unexpected events and interrupts the dorsal system, which is why it is often called a circuit breaker.

The ventral network is strongly right-lateralised, and that asymmetry is the standard explanation for a clinical fact: hemispatial neglect after right-hemisphere damage is common, severe and persistent, while left-hemisphere damage rarely produces its mirror image. Patients with neglect eat from the right of the plate, shave the right of the face, read only the right half of a word, and bisect a line well to the right of its midpoint. Under extinction, they detect a single stimulus on the left perfectly well and miss it when a simultaneous stimulus appears on the right, which is a competition failure, not a sensory one.

Two findings pin down what is preserved. The Bisiach and Luzzatti cathedral result shows the deficit applies to imagined space. And John Marshall and Peter Halligan showed patients two drawings of a house, identical except that one had flames coming from the left-hand window. Patients said the drawings were the same. Asked which house they would rather live in, they consistently chose the one that was not burning, while insisting the choice was arbitrary. The neglected information was processed; it did not reach report.

Posner's own contribution to this literature was to show that right parietal patients have a specific problem: they are disproportionately slow when a cue directs attention to the right and the target then appears on the left. The difficulty is in disengaging from the ipsilesional side, not in seeing the contralesional one.

Common misconceptions

• Neglect is blindness in the left visual field. Visual fields can be intact. The Milan patients described the square with their eyes closed, and extinction appears only when there is competition.
• Neglected information is not processed. The burning house study shows it influences preference while remaining unreportable.
• Attention is a spotlight that sweeps continuously across space. Attention can also select objects and features, and evidence for smooth analogue movement between locations is weak.
• Pop-out means no attention is involved. Efficient search still requires attention to the display and to the task; what pop-out shows is that the target's feature is available without serial checking.
• Illusory conjunctions are guesses. Participants report them with high confidence under conditions of divided attention, which is the point of the demonstration.

Summing up

• Covert attention can be moved without eye movements, producing validity effects of roughly 20 to 50 milliseconds in the Posner task.
• Exogenous orienting is fast and reverses into inhibition of return after about 300 milliseconds; endogenous orienting is slower and sustained.
• Feature search slopes are near flat and conjunction search slopes are around 20 to 30 milliseconds per item, which motivated feature integration theory.
• Illusory conjunctions under divided attention support the claim that attention binds features that are registered separately.
• Duncan and Humphreys reinterpreted search efficiency as a continuum governed by target-distractor and distractor-distractor similarity.
• Moran and Desimone showed attention filters the unattended stimulus out of a V4 receptive field, which biased competition generalises.
• Neglect follows right-hemisphere damage to a right-lateralised ventral attention network, spares sensation, extends to imagined space, and leaves the neglected information able to influence preference.

Sources

1. Moran, J., & Desimone, R. (1985). Selective attention gates visual processing in the extrastriate cortex. Science, 229(4715), 782-784. pubmed.ncbi.nlm.nih.gov
2. Desimone, R., & Duncan, J. (1995). Neural mechanisms of selective visual attention. Annual Review of Neuroscience, 18, 193-222. pubmed.ncbi.nlm.nih.gov
3. Corbetta, M., & Shulman, G. L. (2002). Control of goal-directed and stimulus-driven attention in the brain. Nature Reviews Neuroscience, 3(3), 201-215. pubmed.ncbi.nlm.nih.gov
4. Marshall, J. C., & Halligan, P. W. (1988). Blindsight and insight in visuo-spatial neglect. Nature, 336(6201), 766-767. pubmed.ncbi.nlm.nih.gov
5. Posner, M. I. (1980). Orienting of attention. Quarterly Journal of Experimental Psychology, 32(1), 3-25.
6. Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of attention. Cognitive Psychology, 12(1), 97-136.
7. Bisiach, E., & Luzzatti, C. (1978). Unilateral neglect of representational space. Cortex, 14(1), 129-133.

What was removed, and what it cost

On 1 September 1953 William Scoville drilled two holes above Henry Molaison's eyes, lifted the frontal lobes, and suctioned out the medial surface of both temporal lobes. He recorded the resection as extending eight centimetres back from the temporal poles. The 2014 postmortem reconstruction by Jacopo Annese and colleagues, who cut the frozen brain into 2,401 slices and rebuilt it digitally, showed that the removal was in fact shorter: the amygdala and entorhinal cortex were gone bilaterally, but roughly two centimetres of posterior hippocampus survived on each side, cut off from its normal input. The same study found a small lesion in left orbitofrontal cortex that nobody had known about for sixty years.

Molaison's intelligence was above average and stayed so. His digit span was six. His vocabulary, syntax, perception and personality were unaffected. He could hold a conversation. What he could not do was retain any of it. Turn away for a minute and the conversation was gone, and it stayed gone for fifty-five years.

Remember: The value of this case is not the severity of the loss. It is the sharpness of its edges. A general dimming of memory would have taught us nothing about how memory is organised.

The comparison this lesson is built on

Set out what H.M. could and could not do, and a taxonomy falls out of it.

The vertical line in that table separates declarative memory, which you can bring to mind and state, from nondeclarative memory, which shows up in performance. Neal Cohen and Larry Squire made the distinction rigorous in 1980 with a task that is worth doing yourself: reading text in mirror image. Amnesic patients improved at mirror reading across three days at the same rate as controls, and then failed to recognise the words they had read. Knowing how and knowing that came apart cleanly.

Episodic against semantic

Endel Tulving's division within declarative memory separates remembering an event, with its time, place and the sense of having been there, from knowing a fact stripped of its acquisition context. You know that Oslo is the capital of Norway; you almost certainly cannot recall learning it.

Two patient groups pull these apart. Patient K.C., after a motorcycle accident that damaged the hippocampus bilaterally, retained a vast store of factual knowledge, including facts about his own life, while losing the ability to re-experience any personal event from any period. And in 1997 Faraneh Vargha-Khadem and colleagues described three young people who had suffered hippocampal damage early in life, one at birth. All three had severe episodic memory impairment: they could not report what had happened an hour earlier, or find their way, or remember appointments. All three attended mainstream schools and acquired speech, literacy and factual knowledge within the normal range.

That result is a genuine constraint on theory. Semantic learning must be possible with a compromised hippocampus, which means the routes into the two stores are at least partly separate.

The third system: holding rather than storing

H.M.'s normal digit span shows something the table above compresses: the machinery for holding a small amount of information across seconds is not the machinery for laying anything down. Alan Baddeley and Graham Hitch replaced the old idea of a single short-term store in 1974 with a working memory that has parts.

• The phonological loop. A store that holds speech-based information for a couple of seconds plus a rehearsal process that refreshes it. The evidence is a set of tidy effects: span is worse for lists of long words than short ones, because rehearsal takes longer; span is worse for phonologically similar words (bat, cat, mat) than dissimilar ones; and saying an irrelevant word repeatedly while reading abolishes both effects, because it occupies the rehearsal process.
• The visuospatial sketchpad. The visual and spatial equivalent, which can be loaded independently: a concurrent spatial task interferes with spatial retention while a concurrent verbal task does not, which is a dual-task double dissociation.
• The central executive. Not a store at all but the control system that allocates attention, switches between tasks and coordinates the subsystems.
• The episodic buffer, added in 2000, a limited store where information from the subsystems and from long-term memory is bound into integrated episodes.

The loop looks pointless until you ask what it is for. Giuseppe Vallar and Baddeley's patient P.V. had a specifically impaired phonological store, with a span of two, and normal long-term learning of familiar material. What she could not do was learn the sound patterns of new words in an unfamiliar language. The loop is a device for acquiring new phonological forms, which is a plausible answer to why evolution would build one.

The upshot: Short-term memory is a description of a duration. Working memory is a claim about architecture: separable buffers for different codes plus a control system, all of it testable by finding tasks that interfere with one component and not another.

How much fits? George Miller's magical number seven has been substantially revised. When rehearsal and chunking are prevented, the limit converges on about four items, a figure defended in detail by Nelson Cowan. Chunking is what makes the higher figure attainable: the twelve letters C, I, A, F, B, I, N, A, S, A, N, H, S exceed anyone's span until they are read as four familiar acronyms.

Where does working memory live?

In 1989 Shintaro Funahashi, Charles Bruce and Patricia Goldman-Rakic recorded from monkey dorsolateral prefrontal cortex during an oculomotor delayed response task: a target flashes, then three seconds of darkness, then the animal must look where the target was. They found cells that fired throughout the delay, and fired for one direction of target and not others. These memory fields looked like the substrate of holding something in mind, and persistent prefrontal delay activity became the standard account.

It has been substantially complicated since, in two ways. First, multivariate analyses show that the specific content of what is being held is decodable from sensory cortex, in visual areas for visual material, better than it is from prefrontal cortex, which appears to carry priority and task rules rather than the item. Second, the content can be decoded during delay periods in which no elevated persistent firing is present, so-called activity-silent states, which suggests that short-lived synaptic changes can carry information between moments of active retrieval.

Common misconceptions

• Amnesia means losing your past and your identity. The clinical hallmark is anterograde: an inability to form new declarative memories. Retrograde loss is usually temporally graded, with childhood best preserved.
• H.M. could not learn anything new. He learned motor skills at a normal rate, showed normal priming, and acquired some fragments of post-1953 knowledge. What he could not do was form new episodes and facts he could report.
• Memories are stored in the hippocampus. The hippocampus is required to form new declarative memories and to retrieve recent ones; long-term storage is distributed in neocortex, which is why H.M.'s childhood memories survived its removal.
• Short-term memory is a waiting room that memories pass through on their way to long-term storage. K.F. had a span of two with normal long-term learning, and H.M. had a normal span with no long-term learning, which is a double dissociation against the serial model.
• Working memory holds seven items. With chunking and rehearsal controlled, the limit is closer to four.

Pulling it together

• Scoville's bilateral medial temporal resection left H.M. with above-average intelligence, a normal digit span and no capacity to form new declarative memories.
• The 2014 postmortem reconstruction showed the resection was shorter than recorded, spared about two centimetres of deafferented posterior hippocampus, and included an unknown left orbitofrontal lesion.
• Cohen and Squire showed amnesic patients learn mirror reading at a normal rate while failing to recognise the material, separating knowing how from knowing that.
• Episodic and semantic memory dissociate: developmental amnesic patients with early hippocampal damage acquire school knowledge while failing to remember events.
• Baddeley and Hitch replaced the unitary short-term store with a phonological loop, a visuospatial sketchpad, a central executive and later an episodic buffer.
• The word length, phonological similarity and articulatory suppression effects are the standard evidence for the loop, whose function appears to be learning new phonological forms.
• Prefrontal delay activity carries priority and task structure more than item content, which is decodable from sensory cortex and sometimes from activity-silent states.

Sources

1. Cohen, N. J., & Squire, L. R. (1980). Preserved learning and retention of pattern-analyzing skill in amnesia: Dissociation of knowing how and knowing that. Science, 210(4466), 207-210. pubmed.ncbi.nlm.nih.gov
2. Annese, J., Schenker-Ahmed, N. M., Bartsch, H., Maechler, P., Sheh, C., Thomas, N., et al. (2014). Postmortem examination of patient H.M.'s brain based on histological sectioning and digital 3D reconstruction. Nature Communications, 5, 3122. pubmed.ncbi.nlm.nih.gov
3. Vargha-Khadem, F., Gadian, D. G., Watkins, K. E., Connelly, A., Van Paesschen, W., & Mishkin, M. (1997). Differential effects of early hippocampal pathology on episodic and semantic memory. Science, 277(5324), 376-380. pubmed.ncbi.nlm.nih.gov
4. Funahashi, S., Bruce, C. J., & Goldman-Rakic, P. S. (1989). Mnemonic coding of visual space in the monkey's dorsolateral prefrontal cortex. Journal of Neurophysiology, 61(2), 331-349. pubmed.ncbi.nlm.nih.gov
5. Corkin, S. (2002). What's new with the amnesic patient H.M.? Nature Reviews Neuroscience, 3(2), 153-160. pubmed.ncbi.nlm.nih.gov
6. Baddeley, A. D., & Hitch, G. (1974). Working memory. In G. H. Bower (Ed.), The psychology of learning and motivation (Vol. 8, pp. 47-89). Academic Press.

The gradient, and the argument about it

Theodule Ribot noticed in 1881 that when memory is lost to disease, it goes in a particular order: recent memories first, childhood last. The pattern has held up. H.M. could describe his boyhood in detail and had lost roughly the decade before his surgery. Korsakoff patients show the same shape. A gradient like that is not what you would expect if the medial temporal lobe simply held the memories, and it is the single observation from which the modern argument grows.

The standard consolidation model, developed by Larry Squire and Pablo Alvarez, explains it like this. The hippocampus learns fast and binds together the cortical fragments that make up an experience: the sight, the sound, the place, the words. Over weeks to years, repeated reinstatement of that pattern strengthens direct cortico-cortical connections until the cortex can reinstate the memory without the hippocampus. Consolidation is a transfer of dependence, not of location, and once it is complete the memory survives hippocampal damage. The gradient falls straight out.

Lynn Nadel and Morris Moscovitch attacked that in 1997 with multiple trace theory. Every time you retrieve an episodic memory, they argued, the hippocampus encodes a new trace of that retrieval, bound to its own context. Old memories therefore have many hippocampal traces rather than none, which makes them harder to abolish but never independent. Their prediction is sharp and different: for genuinely episodic memory, rich in specific detail, retrograde amnesia after hippocampal damage should be flat across the lifespan. Only semantic memory, the gist that has been extracted and repeated until it stands alone, becomes hippocampus-independent.

Weighing the evidence

The standard model has the animal data. In contextual fear conditioning, rats given hippocampal lesions one day after training show no memory of the context; rats lesioned a month after training remember it. Similar gradients appear in socially transmitted food preference and trace eyeblink conditioning. Something changes with time, and it changes what the hippocampus is needed for.

Multiple trace theory has the human autobiographical data. When patients with focal hippocampal damage are interviewed with structured instruments that separate internal details, which are episodic and specific to the event, from external details, which are semantic commentary, the loss of internal detail is often flat across all decades of life, while external detail is preserved. Functional imaging in healthy people points the same way: hippocampal activation during vivid recollection does not decline with the age of the memory when the richness of the recollection is controlled.

The current synthesis, trace transformation theory, concedes to both. What changes with time is not merely dependence but the memory itself: a detailed episode is progressively transformed into a schematic, gist-like representation that cortex can support alone, while any surviving detailed version remains hippocampal. This predicts that the gradient you observe depends on which property you measure, which is exactly why two literatures disagreed for twenty years.

Bottom line: Ask not whether a memory is consolidated but what has been consolidated. Gist and detail have different fates and different anatomy.

The mechanism, caught in the act

In 1994 Matthew Wilson and Bruce McNaughton recorded ensembles of hippocampal place cells while rats ran a track, then kept recording while the animals slept. Pairs of cells that had fired together during running fired together again during subsequent slow-wave sleep, far above chance, in sequences that preserved the order of the run and ran roughly twenty times faster than the original experience. Replay occurs in bursts during sharp-wave ripples, brief high-frequency events in which hippocampal output floods cortex.

This is the physiological candidate for what the standard model needs: repeated reinstatement of the cortical pattern, offline, when no new input is competing for the same circuitry.

Jan Born's laboratory turned it into a human experiment. Participants learned card locations in a room scented with rose odour. During subsequent slow-wave sleep the odour was delivered again. Retention improved measurably compared with a night without the cue. The same cue during rapid eye movement sleep did nothing, and the odour during wakefulness did nothing. Targeted memory reactivation has since been replicated many times with sounds as well as odours, and the effect sizes are modest but real.

Key idea: Sleep is not passive for memory. Slow-wave sleep hosts the replay that stabilises declarative memory, and you can bias which memories get replayed by re-presenting a cue that was present at learning.

Retrieval makes a memory fragile again

A consolidated memory ought to be permanent. In 2000 Karim Nader, Glenn Schafe and Joseph LeDoux showed it is not. They trained rats to fear a tone paired with shock. A day later they played the tone, reactivating the memory, and immediately infused anisomycin, a protein synthesis inhibitor, into the amygdala. Tested the next day, those rats showed little fear of the tone. Rats given the same drug without hearing the tone were unaffected, and so were rats given the tone with no drug.

Retrieval had returned the memory to a state that required new protein synthesis to persist, which is what reconsolidation means. The clinical implication was obvious and immediately pursued: reactivate a traumatic memory, interfere with its restabilisation, and weaken it. Merel Kindt and colleagues showed that oral propranolol given with reactivation reduced the later startle response to a conditioned cue in healthy volunteers.

Then the boundary conditions arrived. Reconsolidation is not triggered by every retrieval. It appears to require prediction error, some mismatch between what the memory predicts and what happens, and it is harder to induce for old memories and for strongly trained ones. Translation to clinical PTSD has produced mixed results. The phenomenon is real and the therapeutic story is not settled.

The adult brain still changes shape

Consolidation and reconsolidation are changes in what a circuit holds. There is also change in the circuit itself, and it does not stop at the critical periods of lesson 2.

Michael Merzenich's monkey work in the 1980s showed that somatosensory cortical maps reorganise: amputate a digit and the cortical territory that represented it is taken over within weeks by neighbouring digits; train a monkey to use one fingertip and its representation expands. Vilayanur Ramachandran connected this to the clinic. In some arm amputees, touching the face produces sensation referred to the missing hand, which fits the fact that the hand and face representations are neighbours in the somatosensory map.

In humans, Eleanor Maguire and colleagues scanned London taxi drivers, who spend years memorising 25,000 streets to pass the examination called the Knowledge. Posterior hippocampal grey matter was larger than in controls, anterior smaller, and the posterior volume correlated with years of driving. Because a cross-sectional correlation is compatible with selection, the group later followed trainees prospectively and found the change appeared in those who qualified and not in those who failed or in controls.

The most contested question is whether new neurons are added. In March 2018 Shawn Sorrells and colleagues reported in Nature that neurogenesis in the human dentate gyrus declines sharply through childhood and is undetectable in adults. Three weeks later Maura Boldrini and colleagues reported in another journal that it persists into the eighth decade. Both examined human hippocampal tissue with immunohistochemical markers. The disagreement turns on tissue handling: how quickly the tissue was fixed after death, how long it sat in fixative, and which antibodies were used, all of which affect whether immature neuron markers can still be detected. It is a useful case for a graduate reader, because two competent laboratories with the same nominal question produced opposite answers for reasons that are entirely methodological.

Common misconceptions

• Memories are recordings that get filed away intact. Consolidation transforms them: detail is lost, gist is strengthened, and retrieval can destabilise what is retrieved.
• You can learn new material by playing recordings while you sleep. You cannot. What works is cueing memories you already formed while awake, which biases which of them get replayed.
• Reconsolidation lets you erase memories. It weakens the expression of some memories under specific conditions requiring prediction error, and it works less well the older and stronger the memory.
• The adult brain is fixed. Cortical maps reorganise after amputation and training, and hippocampal grey matter changes measurably in people who acquire a demanding spatial skill.
• Adult human neurogenesis is a settled fact. Two 2018 studies in strong journals reached opposite conclusions, and the discrepancy is driven by tissue fixation and antibody choice rather than by biology.

Looking back

• Ribot's gradient, recent memories lost first, is the observation both consolidation theories are built to explain.
• The standard model treats consolidation as a transfer of dependence from hippocampus to cortex, predicting a steep retrograde gradient.
• Multiple trace theory holds that detailed episodic memory is always hippocampus-dependent, predicting a flat gradient for internal detail and a graded one for semantic content.
• Trace transformation theory reconciles them: gist becomes cortical while any surviving detailed version stays hippocampal, so the gradient depends on what you measure.
• Hippocampal place-cell sequences replay during slow-wave sleep about twenty times compressed, within sharp-wave ripples.
• Re-presenting a learning-associated odour during slow-wave sleep improves declarative retention; the same cue in rapid eye movement sleep or waking does not.
• Reactivation makes a memory labile: protein synthesis inhibition in the amygdala after retrieval abolished conditioned fear, and reconsolidation requires prediction error and weakens with memory age and strength.
• Adult plasticity is structural as well as synaptic, from cortical remapping after amputation to hippocampal change in qualified taxi drivers; adult human neurogenesis remains disputed on methodological grounds.

Sources

1. Nadel, L., & Moscovitch, M. (1997). Memory consolidation, retrograde amnesia and the hippocampal complex. Current Opinion in Neurobiology, 7(2), 217-227. pubmed.ncbi.nlm.nih.gov
2. Wilson, M. A., & McNaughton, B. L. (1994). Reactivation of hippocampal ensemble memories during sleep. Science, 265(5172), 676-679. pubmed.ncbi.nlm.nih.gov
3. Rasch, B., Buchel, C., Gais, S., & Born, J. (2007). Odor cues during slow-wave sleep prompt declarative memory consolidation. Science, 315(5817), 1426-1429. pubmed.ncbi.nlm.nih.gov
4. Nader, K., Schafe, G. E., & Le Doux, J. E. (2000). Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval. Nature, 406(6797), 722-726. pubmed.ncbi.nlm.nih.gov
5. Maguire, E. A., Gadian, D. G., Johnsrude, I. S., Good, C. D., Ashburner, J., Frackowiak, R. S., & Frith, C. D. (2000). Navigation-related structural change in the hippocampi of taxi drivers. Proceedings of the National Academy of Sciences, 97(8), 4398-4403. pubmed.ncbi.nlm.nih.gov
6. Sorrells, S. F., Paredes, M. F., Cebrian-Silla, A., Sandoval, K., Qi, D., Kelley, K. W., et al. (2018). Human hippocampal neurogenesis drops sharply in children to undetectable levels in adults. Nature, 555(7696), 377-381. pubmed.ncbi.nlm.nih.gov
7. Boldrini, M., Fulmore, C. A., Tartt, A. N., Simeon, L. R., Pavlova, I., Poposka, V., et al. (2018). Human hippocampal neurogenesis persists throughout aging. Cell Stem Cell, 22(4), 589-599. pubmed.ncbi.nlm.nih.gov

Fluent, grammatical, and empty

Ask a patient with damage to left posterior temporal cortex what brought them to hospital, and you may get thirty seconds of confident, well-articulated, correctly intonated speech containing almost no information: substitutions of one word for a related one, invented words delivered without hesitation, syntactic frames with the content hollowed out. Ask them to repeat a short phrase and they cannot. Ask them to follow a two-step instruction and they cannot. They are not aware anything is wrong.

Carl Wernicke described this in 1874, in a monograph he wrote at 26, and paired it against the pattern Broca had reported thirteen years earlier: effortful, sparse, agrammatical speech with comprehension largely preserved. Two syndromes, two lesions, and an inference that has shaped every neurology examination since.

The classical model, and how to use it

Wernicke, and then Ludwig Lichtheim in 1885, drew the model as a diagram: a centre for the sound patterns of words in posterior temporal cortex, a centre for the motor patterns of speech in inferior frontal cortex, a pathway connecting them, and links from each to a distributed store of concepts. Norman Geschwind revived it in 1965 and it is still taught, because as a bedside classification it works.

The logic of the two transcortical syndromes is worth pausing on, because it is the model's best moment. If repetition is spared while comprehension or production fails, the perisylvian circuit connecting sound to articulation must be intact and disconnected from the rest of cognition. Patients with transcortical sensory aphasia can echo a long sentence back perfectly and have no idea what it meant.

The point: Repetition is the diagnostic axis that a purely fluency-and-comprehension description misses. Two patients who look alike on the first two dimensions can differ completely on the third, and that difference is anatomical.

Where the diagram is wrong

Three findings have dismantled the specifics while leaving the classification useful.

The lesions are not where the model says. Nina Dronkers and colleagues analysed lesions in a large series of aphasic patients and found that persistent auditory comprehension deficits were associated with damage to left middle temporal gyrus and the white matter beneath it, rather than with the posterior superior temporal gyrus that textbooks label Wernicke's area. Combined with the lesson 1 finding that persistent Broca's aphasia requires damage well beyond Broca's area, the anatomical labels have drifted free of the syndromes named after them.

Nobody agrees where the areas are. Pascale Tremblay and Anthony Dick surveyed researchers in the field and found substantial disagreement about the boundaries of both Broca's and Wernicke's areas, with the latter placed in materially different locations by different respondents. A term used inconsistently by the people who use it most cannot carry a theoretical claim.

Conduction aphasia is not simply a cut cable. The classical account attributes it to severed arcuate fasciculus fibres. Many cases involve cortical damage in the left posterior temporal and supramarginal region, and the deficit behaves more like a failure of an auditory-motor interface than like a broken wire.

The dual-stream model

Gregory Hickok and David Poeppel's replacement borrows the architecture of lesson 6. Auditory input reaches bilateral superior temporal cortex, then divides.

• The ventral stream maps sound onto meaning, running through middle and inferior temporal cortex. Critically, it is bilateral: both hemispheres can do it.
• The dorsal stream maps sound onto articulation, running through a region at the sylvian parietal-temporal junction to frontal motor and premotor cortex. It is strongly left-dominant.

This asymmetry earns its keep by explaining a fact the classical model cannot. Unilateral left temporal damage very rarely abolishes speech comprehension, whereas it routinely devastates production and repetition. If comprehension ran through one left-hemisphere centre, that should not be true. If comprehension is bilateral and the sound-to-articulation route is left-lateralised, it follows directly.

The model also reframes what conduction aphasia is: damage to the auditory-motor interface leaves you able to understand and to speak spontaneously, and unable to convert an incoming sound pattern into an outgoing motor plan, which is exactly the repetition deficit.

Reading, which is too recent for evolution

Writing is about 5,400 years old, and alphabets younger. No cortical region can have been selected for it. Yet in every literate person in every script tested, one patch of left occipitotemporal cortex responds more to written words than to other visual stimuli, in almost the same coordinates. Stanislas Dehaene named it the visual word form area and proposed neuronal recycling: reading colonises cortex evolved for fine-grained shape recognition, and it lands where it does because that region already computes invariance over position and size and sits next to language areas.

Dehaene and colleagues tested this by scanning adults who were illiterate, adults who had learned to read as adults, and adults literate since childhood. Literacy increased the response of that region to written words, sharpened early visual responses, and produced small reductions in the response to faces at nearby sites, which is what a competition-for-cortex account predicts. Learning to read is a measurable structural intervention on the visual system.

The clinical counterpart is old. In 1892 Joseph Dejerine described a man who could write fluently and could not read what he had just written, after a lesion of left occipital cortex plus the splenium of the corpus callosum: visual information reaching only the right hemisphere, with no route to left-hemisphere language.

Dyslexia

Developmental dyslexia affects something like 5 to 10 percent of children depending on the cut-off used, and the best-supported account is a deficit in phonological processing: representing, storing and retrieving the sound structure of words. Dyslexic readers are slower and less accurate at deleting a phoneme from a spoken word, at rapid naming, and at learning arbitrary sound-symbol mappings, and these predict reading outcome better than any visual measure.

Imaging shows reduced engagement of left temporoparietal and occipitotemporal reading circuits, with greater reliance on frontal and right-hemisphere regions. Competing accounts, including a magnocellular visual deficit and a cerebellar automatisation deficit, have not accumulated comparable support.

The strongest evidence that the deficit is phonological rather than orthographic comes from comparing languages. In Italian, where the mapping from letters to sounds is nearly transparent, dyslexic readers read far more accurately than English or French dyslexic readers, but show the same phonological processing deficit and the same reduced activation in left temporo-occipital regions. The underlying impairment is constant; how disabling it is depends on how forgiving the writing system is.

Why this matters: A biological deficit and a cultural artefact interact to produce a disorder. Neither alone predicts who will struggle.

Common misconceptions

• Dyslexia is seeing letters backwards. Letter reversals are common in all beginning readers and are not diagnostic. The core deficit is phonological.
• Broca's area produces speech and Wernicke's area comprehends it. Persistent Broca's aphasia needs damage beyond Broca's area, comprehension deficits track middle temporal damage, and researchers do not agree on where either region is.
• Language is in the left hemisphere. Production and the sound-to-articulation route are strongly left-lateralised in most people, but comprehension is bilateral, and prosody and discourse-level processing draw on the right hemisphere.
• Aphasia is a loss of intelligence. Non-verbal reasoning can be intact, and patients with severe expressive aphasia often follow complex situations perfectly well.
• The brain has a reading module. It has a shape-recognition region that literacy repurposes, at the cost of small reductions in its responses to other categories.

What to remember

• Wernicke's 1874 description paired fluent, paraphasic speech with impaired comprehension against Broca's non-fluent pattern, and Lichtheim and Geschwind turned the pair into a diagram.
• Repetition is the third diagnostic axis, and the transcortical syndromes are the model's best prediction: intact repetition with impaired comprehension or production.
• Lesion mapping places persistent comprehension deficits in left middle temporal cortex and its white matter, not in the posterior superior temporal gyrus the textbooks label.
• Researchers surveyed on the location of Broca's and Wernicke's areas disagree substantially, which limits what those labels can support.
• The dual-stream model divides a bilateral ventral sound-to-meaning route from a left-lateralised dorsal sound-to-articulation route, explaining why left temporal damage spares comprehension more than production.
• The visual word form area is repurposed shape-recognition cortex, and learning to read measurably changes its responses.
• Dyslexia is best explained by a phonological deficit; its severity depends on the transparency of the orthography, which is why Italian dyslexic readers read more accurately than English ones with the same underlying impairment.

Sources

1. Dronkers, N. F., Wilkins, D. P., Van Valin, R. D., Redfern, B. B., & Jaeger, J. J. (2004). Lesion analysis of the brain areas involved in language comprehension. Cognition, 92(1-2), 145-177. pubmed.ncbi.nlm.nih.gov
2. Hickok, G., & Poeppel, D. (2007). The cortical organization of speech processing. Nature Reviews Neuroscience, 8(5), 393-402. pubmed.ncbi.nlm.nih.gov
3. Tremblay, P., & Dick, A. S. (2016). Broca and Wernicke are dead, or moving past the classic model of language neurobiology. Brain and Language, 162, 60-71. pubmed.ncbi.nlm.nih.gov
4. Dehaene, S., Pegado, F., Braga, L. W., Ventura, P., Nunes Filho, G., Jobert, A., et al. (2010). How learning to read changes the cortical networks for vision and language. Science, 330(6009), 1359-1364. pubmed.ncbi.nlm.nih.gov
5. Paulesu, E., Demonet, J. F., Fazio, F., McCrory, E., Chanoine, V., Brunswick, N., et al. (2001). Dyslexia: Cultural diversity and biological unity. Science, 291(5511), 2165-2167.
6. Geschwind, N. (1965). Disconnexion syndromes in animals and man. Brain, 88(2), 237-294 and 88(3), 585-644.

The claim we are going to take apart

Here is the version in most textbooks. On 13 September 1848, near Cavendish, Vermont, a tamping iron blasted through the head of a 25-year-old railway foreman named Phineas Gage. He survived. His intelligence, memory and speech were unimpaired, but his personality was destroyed: previously responsible and well-liked, he became profane, capricious and unemployable, drifting between freak shows until he died. He proves that the frontal lobes are the seat of personality and social judgement.

The physical facts are solid. The iron was 1.1 metres long, 3.2 centimetres in diameter at its widest, and weighed just over 6 kilograms. It entered under the left cheekbone, passed behind the left eye, and exited through the top of the skull, landing some 25 metres away. Gage spoke within minutes and walked to the ox-cart. His physician, John Martyn Harlow, treated him through months of infection, and he lived another twelve years.

The inference is the problem. We are going to trace it step by step and find where it breaks.

Step one: what is the evidence for the personality change?

Two documents. Harlow published a short clinical report in 1848 describing the injury and the recovery, with almost nothing about behaviour. Then, in 1868, twenty years after the accident and eight years after Gage's death, Harlow published a follow-up containing the famous characterisation: fitful, irreverent, profane, impatient of restraint, and no longer Gage.

Notice what that is. It is a single physician's recollection, written two decades later, based on second-hand reports from people who had known Gage, with no systematic observation, no baseline measurement, and no comparison case. It is the entire empirical foundation of the personality claim. Every vivid detail that circulates today has been elaborated from those few sentences.

So what?: A famous case is not the same as a well-documented one. When a finding is repeated for 150 years, the citation trail can be long while the evidence at the bottom of it stays two paragraphs deep.

Step two: what happened to Gage afterwards?

Malcolm Macmillan spent years reconstructing Gage's later life from primary records, and the result does not match the drifter story. Gage went to Chile around 1852 and worked there for roughly seven years as a long-distance stagecoach driver on the route between Valparaiso and Santiago. That work required managing a six-horse team over mountain roads, keeping a schedule, handling fares, and dealing with passengers daily. It is hard to reconcile with an incapacity for planning and social regulation.

Gage returned to his family in San Francisco in 1859 as his health failed, worked on a farm, and died in May 1860 after a series of seizures. The freak-show element of the story rests on a brief appearance with the iron at a museum exhibit and has been inflated far beyond the record.

Macmillan's reading is that Gage showed substantial social recovery, and that the structured, demanding routine of the coach work may have supported it. That reading is itself an inference from thin evidence, and Macmillan says so. The point is not that the opposite story is true. It is that the confident textbook story is unsupported in both directions.

Step three: what was actually damaged?

Gage's skull and the iron are at Harvard. In 1994 Hanna Damasio and colleagues measured the skull, modelled the possible trajectories, and concluded that the rod damaged ventromedial prefrontal cortex bilaterally while sparing regions needed for language and motor control. That reconstruction is the reason Gage is cited as evidence for ventromedial function specifically.

Later reconstructions disagree. Working from computed tomography of the skull rather than from external measurement, subsequent groups argued that the damage was largely confined to the left hemisphere and did not cross the midline as extensively as the 1994 model implied. John Van Horn and colleagues added a connectomic analysis in 2012, estimating that the rod directly damaged something on the order of 4 percent of the cerebral cortex and 11 percent of the white matter, with the fibre-pathway disruption far exceeding the cortical loss.

So the anatomy that the story rests on is itself contested, reconstructed from a skull with holes in it, 150 years after the fact, by methods that give different answers.

Step four: what survives, and what actually established it

Gage remains a legitimate illustration that damage to prefrontal cortex can leave perception, memory, language and intelligence intact while disturbing regulation of behaviour. That is a real and important claim, and it is not established by Gage. It is established by modern patient series in which ventromedial prefrontal damage is documented on imaging, behaviour is measured with instruments, and controls are matched, most influentially by Antonio Damasio's group in Iowa. Lesson 13 takes up what those patients do on a gambling task.

The instruments of control

Executive function is measured, not intuited. Three tasks carry most of the literature.

The Stroop task. John Ridley Stroop reported in 1935 that naming the ink colour of a colour word takes substantially longer when the word and the colour disagree. Say the ink colour of the word RED printed in blue and you will feel the cost, typically 100 milliseconds or more per item. The interference is asymmetric: reading the word is not slowed by the ink colour. Reading is the more practised, more automatic process, and the task requires the weaker one to win.

What resolves the conflict? Matthew Botvinick and colleagues proposed that dorsal anterior cingulate cortex monitors for conflict between competing responses and signals dorsolateral prefrontal cortex to increase control. The theory's best evidence is a behavioural signature: the Stroop effect is smaller on a trial that follows an incongruent trial than on one following a congruent trial, as though the previous conflict had turned control up.

Task switching. Give people a run of trials on one rule, then alternate rules, and responses on switch trials are slower and more error-prone. Stephen Monsell's key observation is that the cost shrinks with preparation time but does not vanish: a residual switch cost survives even when participants know the upcoming task seconds in advance. Part of the cost is the carry-over of the previous task set, which does not fully release until the new task is actually performed.

The Wisconsin card sort. Cards are sorted by colour, form or number; the examiner says only right or wrong; the rule changes without warning after a run of correct sorts. Brenda Milner showed in 1963 that patients with dorsolateral frontal damage achieve few categories and, critically, keep sorting by the rule that has stopped working. Perseveration is the signature: they can often state that the rule has changed and continue sorting by the old one, which separates knowing from doing.

Earl Miller and Jonathan Cohen tied this together in 2001. Prefrontal cortex, on their account, does not execute anything. It represents the current task, holds that representation active against interference, and sends bias signals to the posterior systems that do the work, tipping the competition of lesson 8 toward task-relevant information. Control is a bias, not a controller.

The core of it: Every good measure of executive function creates a situation where the habitual response is wrong. That is the only way to see control at all, because when the habit is correct, control has nothing to do.

Is executive function one thing?

Factor-analytic work by Akira Miyake and colleagues examined performance across many control tasks and found three separable but correlated components: updating the contents of working memory, shifting between task sets, and inhibiting prepotent responses. Later work from the same group suggested that inhibition is not separable from what is common to all three, leaving a common factor plus specific updating and shifting factors. Either way, the finding matters clinically: a patient can fail one and pass another, and a single executive function score conceals that.

Common misconceptions

• Gage proves the frontal lobes are the seat of personality. The behavioural evidence is one physician's recollection written twenty years later, the later-life record shows years of demanding skilled work, and the lesion reconstruction is contested.
• The Stroop effect shows that people cannot inhibit reading. It shows that a stronger automatic process competes with a weaker deliberate one; control resolves the competition rather than switching reading off.
• Perseveration on the card sort means the patient does not understand the task. Many patients state that the rule has changed while continuing to sort by the old one.
• Executive function is a single capacity. Updating, shifting and inhibition are separable enough that a patient can fail one and pass another.
• Prefrontal cortex executes cognitive operations. On the dominant account it maintains task representations and biases posterior systems that do the processing.

The short version

• The entire behavioural record for Gage is Harlow's 1868 recollection, written twenty years after the injury and eight years after Gage's death.
• Macmillan's archival work places Gage in about seven years of skilled stagecoach work in Chile, which is hard to square with the drifter story.
• The 1994 reconstruction placed the damage in bilateral ventromedial prefrontal cortex; later reconstructions argue for largely left-sided damage, with white matter disruption far exceeding cortical loss.
• The claim that prefrontal damage can spare intellect while disturbing behavioural regulation is well supported, but by modern imaged and measured patient series rather than by Gage.
• Stroop interference is asymmetric because reading is the more automatic process, and conflict adaptation is the main evidence for cingulate monitoring driving prefrontal control.
• Switch costs shrink with preparation but leave a residual cost, reflecting carry-over of the previous task set.
• Perseveration on the Wisconsin card sort separates knowing the rule has changed from being able to act on it.
• Miller and Cohen cast prefrontal cortex as a source of bias signals rather than an executor, and factor analyses show executive function is several separable capacities.

Sources

1. Damasio, H., Grabowski, T., Frank, R., Galaburda, A. M., & Damasio, A. R. (1994). The return of Phineas Gage: Clues about the brain from the skull of a famous patient. Science, 264(5162), 1102-1105. pubmed.ncbi.nlm.nih.gov
2. Van Horn, J. D., Irimia, A., Torgerson, C. M., Chambers, M. C., Kikinis, R., & Toga, A. W. (2012). Mapping connectivity damage in the case of Phineas Gage. PLoS ONE, 7(5), e37454. pubmed.ncbi.nlm.nih.gov
3. Monsell, S. (2003). Task switching. Trends in Cognitive Sciences, 7(3), 134-140. pubmed.ncbi.nlm.nih.gov
4. Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167-202. pubmed.ncbi.nlm.nih.gov
5. Macmillan, M. (2000). An odd kind of fame: Stories of Phineas Gage. MIT Press.
6. Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journal of Experimental Psychology, 18(6), 643-662.
7. Milner, B. (1963). Effects of different brain lesions on card sorting: The role of the frontal lobes. Archives of Neurology, 9(1), 90-100.

A woman who could not be frightened, until she was

Patient S.M. has bilateral calcification of the amygdala from a rare genetic condition, Urbach-Wiethe disease, and has been studied for three decades. She cannot recognise fear in a photographed face, though she recognises happiness and sadness. Taken to an exotic pet shop she handled snakes she had said she would avoid, and had to be stopped from touching a tarantula. In a haunted-house attraction she led the group and startled the actors. Asked to rate her fear during horror films, she reported almost none.

In 2013 Justin Feinstein, Colin Buzza, Ralph Adolphs and colleagues had S.M. and two other patients with bilateral amygdala damage inhale a single breath of air containing 35 percent carbon dioxide. All three had full-blown panic attacks, with terror they described as unlike anything they could remember. Healthy controls panicked less often.

Remember: The structure whose loss removes fear of snakes, faces and haunted houses does not remove fear of suffocation. Whatever the amygdala does, it is not the place where the feeling of fear is generated.

Position one: a defensive survival circuit

The best-characterised circuit in affective neuroscience is Joseph LeDoux's auditory fear conditioning pathway in the rat. Pair a tone with a footshock and the animal freezes to the tone thereafter. Trace the anatomy and you find the following.

• Auditory information reaches the lateral amygdala by two routes: directly from auditory thalamus, fast and coarse, and via auditory cortex, slower and detailed. LeDoux called these the low road and the high road.
• The lateral amygdala is where the tone and the shock converge, and where synaptic strengthening occurs.
• The central nucleus is the output. Its projections to the periaqueductal grey produce freezing, to the lateral hypothalamus produce blood pressure changes, and to hypothalamic nuclei produce stress hormone release. Lesion one output and you remove one component of the response while the others remain.

Extinction adds a second circuit. Present the tone repeatedly without shock and freezing declines, but the original association is not erased: it returns if the animal is tested in a different context (renewal), after an unsignalled shock (reinstatement), or simply after time (spontaneous recovery). Extinction is new inhibitory learning, driven by ventromedial prefrontal cortex acting on the amygdala, layered on top of a memory that persists. This is the mechanistic basis of exposure therapy and of relapse after it.

Two qualifications belong with this account. The low road is solidly established in rodents; the case for a fast subcortical route to the human amygdala is much weaker than popular accounts suggest, and has been challenged on both anatomical and functional grounds. And LeDoux himself now argues that the whole circuit should not be called a fear system at all. On his current view it detects threat and organises defensive responses, and does so whether or not anything is felt; the conscious experience of fear is a separate construction that depends on cortical networks and on having a concept of fear. He has spent years trying to persuade the field to stop using the word for the circuit, because the slippage between the two meanings has caused real confusion in translational psychiatry.

Position two: the body as a guide to decisions

Antonio Damasio's somatic marker hypothesis holds that decisions are guided by bodily states associated with the outcomes of past choices, represented in ventromedial prefrontal cortex, and that these markers bias choice before and beneath conscious reasoning.

The evidence is the Iowa gambling task. Four decks: two pay 100 dollars a card with occasional heavy penalties and lose money over time, two pay 50 dollars with small penalties and gain money over time. Healthy participants drift toward the good decks. Patients with ventromedial prefrontal damage keep returning to the bad ones, sometimes to bankruptcy, and can describe the rules perfectly well when asked afterwards. Antoine Bechara and colleagues added skin conductance: healthy participants developed an anticipatory skin conductance response before reaching toward a bad deck, and did so, in the original report, before they could verbalise which decks were bad. The patients never developed it.

That result was widely read as showing that a bodily signal guides good choices before conscious knowledge arrives. Two critiques have weakened it substantially.

Tiago Maia and James McClelland re-ran the task with a much finer questionnaire, asking after every ten cards how much participants would win or lose from each deck rather than asking a coarse question about which decks were risky. Participants turned out to have far more explicit knowledge, far earlier, than the original probes had detected, and that knowledge was sufficient to account for their choices. The non-conscious guidance the hypothesis needs was an artefact of under-measuring awareness.

Barnaby Dunn, Tim Dalgleish and Andrew Lawrence then evaluated the hypothesis in detail and found the supporting evidence weak at several joints: anticipatory skin conductance findings replicate inconsistently, the gambling task is not a pure measure of anything and confounds reversal learning with risk and with working memory, and the proposed link from interoception to choice quality was largely untested.

What matters here: The clinical observation is robust. Ventromedial prefrontal patients make disastrous real-world decisions while reasoning normally in the abstract. The mechanistic story about non-conscious bodily markers is the part that did not survive close measurement.

Position three: emotions are constructed, not located

A basic-emotion account holds that a small number of emotions, fear, anger, disgust, sadness, happiness, surprise, are biologically given, with characteristic expressions, physiology and neural circuits. If that is right, imaging should find consistent, category-specific neural signatures.

Kristen Lindquist, Tor Wager, Lisa Feldman Barrett and colleagues meta-analysed the imaging literature and found no such one-to-one mapping. The amygdala was engaged by fear and also by anger, disgust, salience and novelty; the insula by disgust and also by interoception generally. What the data showed was a set of domain-general networks, for interoception and salience, for conceptual knowledge, and for executive control, combining in different weightings.

Barrett's theory of constructed emotion builds on that: an instance of fear is a categorisation of interoceptive and sensory input using learned concepts, in the same way a percept is a categorisation of visual input. Emotion categories are real as categories, and they are not natural kinds implemented in dedicated circuits.

The counter-argument is not weak. Conserved subcortical circuits for defence, care and play are found across mammals; some facial expressions are recognised above chance across cultures; and multivariate analyses since the meta-analysis have shown that emotion categories can be decoded from distributed patterns of brain activity with reasonable accuracy. That last point is important: the failure to find a fear region does not mean there is no reliable fear signature, only that it is not a place.

What would move the argument

Three things, none of them a single decisive experiment. First, measures of subjective experience that do not simply ask participants to choose an emotion word, since the constructionist account predicts that the words themselves shape the report. Second, tests of whether the multivariate signatures of emotion categories generalise across people, across induction methods, and across cultures, which distinguishes a biologically given category from a learned one. Third, careful separation, in patients and in animals, of defensive responding from reported feeling, which is precisely what the carbon dioxide result in S.M. did: response and feeling came apart, and only a theory that separates them survives.

Common misconceptions

• The amygdala is the fear centre. Bilateral damage abolishes fear of snakes and of fearful faces and leaves panic to carbon dioxide intact, and the same structure responds to salience, novelty and reward.
• Extinction erases a fear memory. Renewal, reinstatement and spontaneous recovery all show the original association persists under new inhibitory learning, which is why exposure therapy relapses.
• The low road is a well-established human pathway. It is solid in rodents; the human evidence for a fast subcortical route to the amygdala is contested.
• The Iowa gambling task shows the body knowing before the mind. With finer questioning, participants proved to have explicit knowledge early enough to explain their choices.
• Failing to find a fear region means fear is not real in the brain. Distributed multivariate signatures can be reliable even when no single region is specific.

Where this leaves us

• Patient S.M., with bilateral amygdala damage, shows no fear of snakes, horror films or haunted houses, and panics on inhaling carbon dioxide.
• The rodent fear conditioning circuit runs from auditory thalamus and cortex to lateral amygdala, with central nucleus outputs producing separable components of the defensive response.
• Extinction is new prefrontal-driven inhibitory learning rather than erasure, which explains relapse after exposure therapy.
• LeDoux now argues the circuit implements threat detection and defence, and that conscious fear is a separate cortical construction.
• The somatic marker hypothesis rests on the Iowa gambling task, where ventromedial patients persist with disadvantageous decks and lack anticipatory skin conductance responses.
• Finer probes of awareness showed participants knew more, sooner, than the original studies detected, and a systematic evaluation found the supporting evidence weak at several points.
• Meta-analysis found no one-to-one mapping between discrete emotions and brain regions, supporting a constructionist account, while distributed multivariate signatures for emotion categories do exist.

Sources

1. Feinstein, J. S., Buzza, C., Hurlemann, R., Follmer, R. L., Dahdaleh, N. S., Coryell, W. H., et al. (2013). Fear and panic in humans with bilateral amygdala damage. Nature Neuroscience, 16(3), 270-272. pubmed.ncbi.nlm.nih.gov
2. Phelps, E. A., & LeDoux, J. E. (2005). Contributions of the amygdala to emotion processing: From animal models to human behavior. Neuron, 48(2), 175-187. pubmed.ncbi.nlm.nih.gov
3. LeDoux, J. E. (2014). Coming to terms with fear. Proceedings of the National Academy of Sciences, 111(8), 2871-2878. pubmed.ncbi.nlm.nih.gov
4. Bechara, A., Damasio, A. R., Damasio, H., & Anderson, S. W. (1994). Insensitivity to future consequences following damage to human prefrontal cortex. Cognition, 50(1-3), 7-15. pubmed.ncbi.nlm.nih.gov
5. Maia, T. V., & McClelland, J. L. (2004). A reexamination of the evidence for the somatic marker hypothesis: What participants really know in the Iowa gambling task. Proceedings of the National Academy of Sciences, 101(45), 16075-16080. pubmed.ncbi.nlm.nih.gov
6. Dunn, B. D., Dalgleish, T., & Lawrence, A. D. (2006). The somatic marker hypothesis: A critical evaluation. Neuroscience and Biobehavioral Reviews, 30(2), 239-271. pubmed.ncbi.nlm.nih.gov
7. Lindquist, K. A., Wager, T. D., Kober, H., Bliss-Moreau, E., & Barrett, L. F. (2012). The brain basis of emotion: A meta-analytic review. Behavioral and Brain Sciences, 35(3), 121-143. pubmed.ncbi.nlm.nih.gov

A neuron that stops responding to the thing it likes

A thirsty monkey sits with a recording electrode in the midbrain, in the ventral tegmental area. A drop of apple juice arrives unannounced, and a dopamine neuron that had been ticking along at three or four spikes per second fires a short burst.

Now train the animal for a few days so that a light always precedes the juice by two seconds. Record again. The burst at the juice has gone. The same neuron now bursts to the light. Nothing has changed about the juice, the monkey's thirst, or how much the monkey drinks. What changed is that the juice became predictable.

Then withhold the juice on a probe trial, after the light. The neuron bursts to the light as usual, and then, at precisely the moment the juice was due, its firing drops below baseline for a few hundred milliseconds and recovers.

Key idea: A response that vanishes when the event becomes predictable, and goes negative when a predicted event fails to occur, is not a report of reward. It is a report of the difference between what was expected and what happened.

The procedure, and the arithmetic

Wolfram Schultz, Peter Dayan and Read Montague made that identification precise in 1997 by mapping the recordings onto temporal-difference learning, an algorithm from machine learning. Define a value function V(s) as the total future reward expected from state s. At each step, compute a prediction error: the reward actually received, plus the value of the state you have arrived in, minus the value of the state you left. In symbols, with a discount factor gamma, delta equals r plus gamma times V of the new state, minus V of the old state. Then use delta to update V. That is the whole algorithm.

Run the three experimental conditions through it.

Three points are worth drawing out. The transfer of the burst from reward to cue is the algorithm's signature, because value propagates backward in time to the earliest reliable predictor. The dip requires precise timing, since the neuron must know when the reward was due, which means the system carries a temporal expectation and not just a magnitude. And the fact that the same neurons respond to a cue that predicts juice and to unexpected juice, but not to predictable juice, rules out a simple hedonic reading.

Changing one input flips the result

Repeat the experiment with a cue that predicts juice only half the time. Now the cue raises expected value by half the maximum, so the burst at the cue is smaller. When the juice arrives, it exceeds the prediction, so there is a positive error and a burst; when it does not, there is a negative error and a dip. Averaged over trials the reward-time response is near zero, but on individual trials it is systematically positive or negative, which is exactly what the algorithm predicts and what a reward-magnitude account does not.

Scale it up and the same variable predicts human behaviour. In imaging studies, striatal responses track model-derived prediction errors trial by trial; in Parkinson's disease, where the substantia nigra pars compacta degenerates, patients off medication learn better from negative outcomes than positive ones, and the pattern reverses on dopaminergic medication.

The causal step

All of the above is correlational, which lesson 5 warned about. The causal test came from Elizabeth Steinberg, Patricia Janak and colleagues in 2013, using the blocking paradigm. If a cue already fully predicts a reward, and you add a second cue alongside it, animals do not learn about the new cue: there is no prediction error, so there is nothing to learn from. Steinberg and colleagues optogenetically stimulated ventral tegmental area dopamine neurons at the moment of the reward during those compound trials, manufacturing a prediction error where the animal's own system had none. The animals then learned about the redundant cue.

The upshot: A dopamine burst does not accompany learning; it causes it. Generating one artificially at the right moment produces learning that would otherwise not occur.

Dopamine is not pleasure

The popular identification of dopamine with pleasure survives in headlines and nowhere else. Kent Berridge and Terry Robinson separated two components of reward that everyday language runs together.

• Liking is the hedonic reaction itself, measurable in rats and human infants as a stereotyped pattern of orofacial responses to sweet versus bitter tastes. It depends on opioid and endocannabinoid signalling in small hedonic hotspots in the nucleus accumbens and ventral pallidum.
• Wanting is incentive salience: the property that makes a cue grab attention and motivate approach and effort. This is what dopamine carries.

Dopamine-depleted rats will not work for food and can starve in its presence, yet their orofacial liking reactions to sucrose are entirely normal. Sensitising the dopamine system amplifies wanting without amplifying liking. Robinson and Berridge's incentive sensitisation account of addiction rests on exactly this dissociation: what escalates in addiction is craving triggered by cues, while the pleasure of the drug often declines.

The clinical evidence points the same way. Dopamine agonist medication for Parkinson's disease produces impulse control disorders, including pathological gambling, compulsive shopping and hypersexuality, in a substantial minority of treated patients, and these typically resolve when the drug is reduced. That is a disorder of wanting.

Where the account is incomplete

Four qualifications, each supported and each limiting.

1. Dopamine neurons are not one population. A subset in the dorsolateral substantia nigra is excited by aversive and salient events rather than showing the value-signed response, which suggests parallel channels for value and for salience.
2. Dopamine also signals vigour. Slower, ramping dopamine signals in striatum track the value of the current situation and the invigoration of ongoing movement, which is not a prediction error at all.
3. Prediction error learning is model-free. It caches values without representing the structure of the world, so it cannot explain the flexible re-planning that people show when a goal is devalued. Two-step tasks separate model-free from model-based control, and human behaviour is a mixture of both.
4. Value has to be compared somewhere. Camillo Padoa-Schioppa and John Assad found neurons in orbitofrontal cortex encoding the value of an offered good in a way that does not depend on which good it is, or on what else is on the menu, which is what a common currency for choice requires.

Common misconceptions

• Dopamine is the pleasure chemical. Dopamine-depleted rats show normal hedonic reactions to sweetness and will not work for food. Pleasure is opioid and endocannabinoid signalling in small hotspots; dopamine carries wanting.
• Dopamine neurons signal how rewarding something is. They signal the difference between expected and received value, which is why a fully predicted reward elicits nothing.
• Addiction is a flood of dopamine producing extreme pleasure. On the incentive sensitisation account, cue-triggered wanting escalates while liking often declines.
• The brain has a reward centre. Reward involves midbrain dopamine neurons, striatal targets, hedonic hotspots, orbitofrontal value coding and ventromedial prefrontal comparison, and these components dissociate.
• Prediction errors explain all learning. They are model-free. Flexible re-planning after a goal changes requires a model of the world that this mechanism does not build.

Recap

• Midbrain dopamine neurons burst to unpredicted reward, transfer that burst to a reliable predictive cue, and dip below baseline when a predicted reward is omitted.
• These three patterns match the temporal-difference prediction error, delta equals reward plus discounted new-state value minus old-state value.
• The dip at the expected reward time shows the system carries a temporal expectation, not just an expected magnitude.
• Optogenetically manufacturing a dopamine burst during blocking causes learning about a cue that would otherwise be ignored, which makes the signal causal rather than correlated.
• Wanting and liking dissociate: dopamine depletion abolishes effort for food while leaving hedonic reactions intact.
• Impulse control disorders under dopamine agonist treatment are disorders of wanting and remit when the drug is reduced.
• The account is incomplete: dopamine neurons are heterogeneous, ramping signals track vigour, model-free learning cannot explain goal devaluation effects, and orbitofrontal neurons supply the menu-invariant value signal choice requires.

Sources

1. Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593-1599. pubmed.ncbi.nlm.nih.gov
2. Steinberg, E. E., Keiflin, R., Boivin, J. R., Witten, I. B., Deisseroth, K., & Janak, P. H. (2013). A causal link between prediction errors, dopamine neurons and learning. Nature Neuroscience, 16(7), 966-973. pubmed.ncbi.nlm.nih.gov
3. Berridge, K. C., & Robinson, T. E. (2009). Dissecting components of reward: Liking, wanting, and learning. Current Opinion in Pharmacology, 9(1), 65-73. pubmed.ncbi.nlm.nih.gov
4. Padoa-Schioppa, C., & Assad, J. A. (2006). Neurons in the orbitofrontal cortex encode economic value. Nature, 441(7090), 223-226. pubmed.ncbi.nlm.nih.gov
5. Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.), Chapter 6: Temporal-difference learning. MIT Press.

Four ways to infer a mind from a measurement

This lesson compares four literatures that share a problem. In each, a neural measurement is taken to license a claim about mental states, either someone else's or the participant's own. In each, the measurement is real and the inference is where the arguing happens.

Parma, about 1990

In Giacomo Rizzolatti's laboratory, electrodes in monkey premotor area F5 were recording cells that fire when the animal grasps an object. Between trials, an experimenter reached for something in front of the animal, and one of these motor cells fired while the monkey sat still and watched. Roughly 10 to 20 percent of F5 neurons turned out to have this property, and similar cells were found in inferior parietal cortex. Some were strictly congruent, firing for the same grip observed and executed; most were broadly congruent, matching at the level of the goal.

The interpretive claim followed quickly and grew: mirror neurons constitute a mechanism for understanding the actions of others by simulating them in one's own motor system, and, by extension, a basis for imitation, empathy, the evolution of language, and, in their dysfunction, autism.

Gregory Hickok's 2009 critique set out the problems, and they are serious.

• Monkeys with F5 damage still recognise and respond appropriately to observed actions, which is what a necessity claim would have to rule out.
• People understand actions they cannot perform, including actions no human body can perform, and patients with severe apraxia who cannot execute an action often still recognise it.
• The proposed extension to speech perception fails on the same evidence: damage to motor speech areas can leave speech comprehension largely intact, which lesson 11 established.
• The broken-mirror account of autism has not accumulated the predicted evidence, and imitation deficits in autism are inconsistent.
• Mirror properties may be learned. Cecilia Heyes has argued that sensorimotor experience, watching your own hand grasp thousands of times, is sufficient to produce the association, which fits evidence that mirror responses can be trained and even reversed.

Direct human evidence arrived in 2010, when Roy Mukamel, Itzhak Fried and colleagues recorded single neurons in epilepsy patients during both execution and observation of actions and found cells responding to both. Notably, these were concentrated in supplementary motor area and medial temporal lobe rather than in the classical mirror regions, and a substantial number showed the opposite pattern, excited during execution and inhibited during observation, which fits a role in distinguishing self from other as much as in simulating others.

In short: Mirror neurons exist. The claim that they are the mechanism of action understanding does not follow from their existence and has not survived its own predictions.

Attributing a belief

A different route to other minds is theory of mind: representing what someone else believes, including when it is false. The behavioural test is the false belief task. Sally puts a marble in a basket and leaves; Anne moves it to a box; where will Sally look? Typically developing children answer correctly from around four, having previously answered with their own knowledge rather than Sally's.

Rebecca Saxe and Nancy Kanwisher identified a network that responds when people read about beliefs: bilateral temporoparietal junction, strongest on the right, plus medial prefrontal cortex and precuneus. The selectivity is impressive. The right temporoparietal junction responds more to a story about what someone thinks than to a story about their appearance, their body, or a false photograph, which controls for representing an outdated state of affairs.

Now the caution, which this course has been building toward. The right temporoparietal junction is also one of the two hubs of the ventral attention network from lesson 8, where it responds to unexpected relevant events with no social content whatsoever. So activation there does not by itself indicate mental state attribution. Reading a claim in that direction requires either selectivity established in the same participants, or evidence from a different method. Lesson 17 formalises exactly this error.

Seeing without seeing

Lawrence Weiskrantz and Elizabeth Warrington's patient D.B. had part of his occipital cortex removed to treat a vascular malformation, leaving a blind region in the opposite visual field. He reported seeing nothing there. Asked to guess whether a spot had appeared above or below a line, or to reach toward a target he denied seeing, he performed far above chance. Later patients showed above-chance discrimination of motion direction and of some shapes.

Blindsight is usually attributed to pathways that bypass V1: retinal projections to the superior colliculus and pulvinar, reaching area MT directly. Two cautions belong with it. First, spared islands of V1 tissue account for some cases, so the anatomy must be shown, not assumed. Second, there is a measurement problem. A patient who has a faint degraded experience but applies a strict criterion for saying they saw something will report blindness while performing above chance, which looks like unconscious vision and is not. Careful work has addressed this with confidence ratings and signal detection analysis, and the phenomenon survives in some patients, but the criterion objection is why blindsight cannot be used casually as proof of unconscious perception.

Two hemispheres, one person?

To control intractable epilepsy, a small number of patients in the 1960s had the corpus callosum severed. Roger Sperry and Michael Gazzaniga tested them with stimuli flashed briefly to one visual field, too fast for an eye movement, so that the information reached only one hemisphere.

The results are famous. A word flashed to the left visual field, reaching the right hemisphere, could not be named, because speech is left-lateralised, yet the left hand could select the corresponding object from a hidden array. The most instructive experiment is Gazzaniga's chicken claw study: a chicken claw is shown to the left hemisphere and a snow scene to the right, and the patient is asked to choose related pictures from an array with both hands. The right hand picks a chicken; the left hand picks a shovel. Asked why, the patient, answering from the talking left hemisphere, explains that a shovel is needed to clean out the chicken shed. The explanation is confident, immediate and fabricated. Gazzaniga called this the interpreter: a left-hemisphere process that constructs a coherent account of behaviour from whatever information it has, including none.

The recent challenge is worth taking seriously. Yair Pinto, Victor Lamme and colleagues tested two split-brain patients and found that both could accurately report the presence and location of stimuli anywhere in the visual field, with either hand, verbally or by pointing, even though they could not match a stimulus in one field to a stimulus in the other. Their conclusion is that perception is split while conscious awareness remains unified. The finding is from two patients and is contested. Its value here is that a textbook staple, taught as settled for fifty years, turns out to depend on which responses were tested.

Worth holding on to: The split-brain literature does not show two people in one head. It shows that specific kinds of information fail to cross, and that the verbal hemisphere will explain the resulting behaviour anyway.

What a correlate can show

The standard method for finding a neural correlate of consciousness is contrastive: hold the stimulus constant and compare trials where the participant reports awareness with trials where they do not, using binocular rivalry, masking or attentional blink. The persistent confound is report itself. Pressing a button involves attention, decision and motor preparation, so the contrast includes the consequences of being conscious rather than only its substrate. No-report paradigms, which infer awareness from eye movements or pupil responses, were designed to strip that out, and they shrink the frontal contribution considerably while leaving a posterior contribution.

Christof Koch, Marcello Massimini, Melanie Boly and Giulio Tononi's review sets out what remains: a posterior hot zone in parietal-occipital-temporal cortex whose activity tracks the content of experience more closely than frontal activity does, and an unresolved argument between theories, including global workspace accounts and integrated information theory. A correlate, however reliable, does not distinguish the mechanism of experience from its prerequisites or its consequences. That is not a defect of the experiments; it is the shape of the problem.

Common misconceptions

• People are left-brained or right-brained. Lateralisation is real for specific functions such as speech production, and it does not sort people into analytic and creative types. Large imaging studies find no such division of individuals.
• Mirror neurons explain empathy and autism. Both claims outran the evidence: F5 lesions do not abolish action recognition, and the broken-mirror account of autism has not been supported.
• Blindsight proves unconscious vision. It survives careful testing in some patients, but degraded conscious vision reported under a strict criterion produces the same pattern, so the analysis matters.
• Split-brain patients contain two conscious people. Recent testing found unified awareness of stimulus presence and location with split perception, and the classic demonstrations depend heavily on which responses are permitted.
• Right temporoparietal activation means mental state attribution. The same region is a hub of the ventral attention network and responds to unexpected events with no social content.

What to carry forward

• Mirror neurons are premotor and parietal cells that fire during both execution and observation of goal-directed actions, mostly congruent at the level of the goal.
• Action-understanding claims fail several tests: F5 lesions spare action recognition, people understand actions they cannot perform, and mirror properties may be learned through sensorimotor experience.
• Human single-neuron recordings found execution-observation cells in supplementary motor and medial temporal regions, many of them inhibited during observation.
• The theory of mind network centres on right temporoparietal junction and medial prefrontal cortex, and the same temporoparietal region belongs to the ventral attention network, which is a standing reverse-inference trap.
• Blindsight is above-chance forced-choice performance in a field reported as blind, attributed to collicular and pulvinar routes, with spared V1 islands and response criterion as the two competing explanations.
• Split-brain testing shows information failing to cross and a left-hemisphere interpreter that confabulates explanations; recent work in two patients found unified awareness with split perception.
• Contrastive methods for neural correlates of consciousness are confounded by report, and no-report paradigms reduce the frontal contribution while leaving a posterior one.

Sources

1. Hickok, G. (2009). Eight problems for the mirror neuron theory of action understanding in monkeys and humans. Journal of Cognitive Neuroscience, 21(7), 1229-1243. pubmed.ncbi.nlm.nih.gov
2. Mukamel, R., Ekstrom, A. D., Kaplan, J., Iacoboni, M., & Fried, I. (2010). Single-neuron responses in humans during execution and observation of actions. Current Biology, 20(8), 750-756. pubmed.ncbi.nlm.nih.gov
3. Weiskrantz, L., Warrington, E. K., Sanders, M. D., & Marshall, J. (1974). Visual capacity in the hemianopic field following a restricted occipital ablation. Brain, 97(4), 709-728. pubmed.ncbi.nlm.nih.gov
4. Pinto, Y., Neville, D. A., Otten, M., Corballis, P. M., Lamme, V. A., de Haan, E. H., et al. (2017). Split brain: Divided perception but undivided consciousness. Brain, 140(5), 1231-1237. pubmed.ncbi.nlm.nih.gov
5. Koch, C., Massimini, M., Boly, M., & Tononi, G. (2016). Neural correlates of consciousness: Progress and problems. Nature Reviews Neuroscience, 17(5), 307-321. pubmed.ncbi.nlm.nih.gov
6. Saxe, R., & Kanwisher, N. (2003). People thinking about thinking people: The role of the temporo-parietal junction in theory of mind. NeuroImage, 19(4), 1835-1842.
7. di Pellegrino, G., Fadiga, L., Fogassi, L., Gallese, V., & Rizzolatti, G. (1992). Understanding motor events: A neurophysiological study. Experimental Brain Research, 91(1), 176-180.

An arithmetic argument that a graduate student made in public

In late 2008 a manuscript began circulating with the title Voodoo Correlations in Social Neuroscience. Ed Vul, then a graduate student, with Christine Harris, Piotr Winkielman and Harold Pashler, had surveyed 55 published papers reporting correlations between fMRI activity and personality or emotion measures, and written to the authors to ask how they had done the analysis. Over half had done it the same way: run a whole-brain correlation, find the voxels where the correlation with the behavioural measure was strongest, then report the correlation in those voxels.

The reason that is fatal takes one line of arithmetic. The highest correlation two measures can show is limited by their reliabilities: the ceiling is the square root of the product of the two. Test-retest reliability for fMRI activation is around 0.7 at best; for a good personality questionnaire, around 0.8. The ceiling is therefore about 0.74. The surveyed papers reported a median correlation near 0.8, with values of 0.85 and 0.88 not unusual, which is above the ceiling for the measures being correlated.

Those values were not measurements. They were the maximum of a large set of noisy correlations, selected for being maximal, and then reported as though the selection had not happened. This is the circular analysis of lesson 4, applied to individual differences.

Bottom line: When an effect size exceeds what the reliability of your instruments permits, the number is telling you about your analysis, not about brains.

The dispute, stated fairly

The paper was published, after argument, under the milder title Puzzlingly High Correlations, alongside commentaries and a reply. Two positions emerged, and both have serious people behind them.

The evidence that accumulated

Power. In 2013 Katherine Button, John Ioannidis, Marcus Munafo and colleagues estimated the statistical power of studies in 49 neuroscience meta-analyses and found a median around 21 percent. Three consequences follow mechanically. A low-powered study is unlikely to detect a real effect. When it does report one, the probability that the effect is real is lower than the p value suggests, because the ratio of true to false positives depends on power and on the prior odds. And any effect that clears significance in a small sample must be large, so published effect sizes are systematically inflated, which is the winner's curse.

Statistical validity. Eklund and colleagues' cluster-inference result from lesson 4 belongs here: a widely used correction was empirically invalid at lenient thresholds for fifteen years.

Analytic flexibility. In 2020 Rotem Botvinik-Nezer and colleagues gave the same fMRI dataset to 70 independent analysis teams with nine pre-specified hypotheses. No two teams used identical pipelines. For five of the nine hypotheses, the teams split on the binary conclusion, in some cases close to evenly. The data were fixed; the answer depended on who analysed them.

Sample size for individual differences. In 2022 Scott Marek, Brenden Tervo-Clemmens, Nico Dosenbach and colleagues used three large datasets, together comprising tens of thousands of scans, to estimate the true size of brain-wide associations between imaging measures and behavioural or cognitive traits. The typical effect is very small. At the sample sizes common in the literature, around 25 participants, the correlations that reach significance are almost entirely noise, and replicating a brain-wide association reliably requires thousands of participants.

Marek's result explains Vul's. If the true correlation between a voxel's activity and a personality score is near 0.1, and you have 20 participants and 100,000 voxels, then the largest observed correlation will be around 0.8 by chance alone. Both papers describe the same phenomenon from opposite ends.

Why this matters: The problem was never that researchers were dishonest. It was that a standard sample size, a standard pipeline and a standard reporting convention combined to produce results that could not have been right.

What the critiques do not show

Three limits are worth stating precisely, because the sceptical case is often overstated in the other direction.

• Not all effects are individual differences. Within-subject contrasts with large effects, such as the response of visual cortex to a flickering checkerboard, or the N400 to a semantic anomaly, are detectable in a handful of participants and replicate across decades and continents. The replication problem concentrates in correlations between a brain measure and a trait measured once per person.
• A failed replication is not proof of a false original. It shifts the evidence, and its own power, fidelity and context matter. In the Open Science Collaboration's replication of 100 psychology studies, 36 percent of replications reached significance against 97 percent of originals, and effect sizes were about half; the correct reading is that a substantial fraction of the original literature is unreliable, not that any particular study is refuted.
• The critiques were made from inside. Nichols, Poldrack, Munafo and others who documented these problems are neuroimagers who then built the fixes. The field caught this, which is what the field is supposed to do.

What actually changed

1. Preregistration and registered reports. Specifying hypotheses, exclusion rules and analysis before seeing the data removes the flexibility that generates false positives; registered reports go further, with peer review and acceptance in principle before data collection, so results cannot influence publication.
2. Permutation testing as the default. Directly answering the cluster-inference failure.
3. Data and code sharing. Repositories for raw datasets and for unthresholded statistical maps mean an analysis can be re-run and a map can be inspected rather than summarised by a table of peak coordinates.
4. Reporting standards. Community guidelines specify what must be reported about acquisition, preprocessing and inference, which makes pipeline choices visible.
5. Consortium-scale samples. Datasets with thousands to hundreds of thousands of participants make brain-behaviour questions answerable at the effect sizes that actually exist.
6. Out-of-sample prediction. Reporting how well a model predicts held-out participants rather than how well it fits the sample it was built on, which makes circularity impossible by construction.

How to read a paper after all this

Four questions, in order. What is the sample size, and is the effect a within-subject contrast or a between-subject correlation? Were the regions of interest defined independently of the reported test? Was the inference corrected, and by what method? And is the effect size plausible given the reliability of the two measures? A paper that answers all four well is worth your attention regardless of how surprising its claim is; a paper that fails the fourth is not worth reading further.

Common misconceptions

• The replication crisis shows that fMRI does not work. It shows that a particular class of design, small-sample brain-behaviour correlation, was uninformative. Within-subject effects with large true sizes replicate reliably.
• The problem was p-hacking by dishonest researchers. The dominant mechanism was standard practice: normal sample sizes, ordinary pipeline flexibility, and a convention of selecting and reporting peaks.
• A large correlation is strong evidence. If it exceeds the ceiling set by the reliabilities of the two measures, it is evidence of a selection error.
• Preregistration prevents exploration. It separates exploration from confirmation, and preregistered reports routinely include labelled exploratory analyses.
• A failed replication refutes the original finding. It updates the evidence, weighted by its own power and fidelity.

The takeaway

• Over half of surveyed brain-personality papers selected voxels by the correlation they then reported, producing values above the ceiling set by measurement reliability.
• The reliability ceiling is the square root of the product of the two reliabilities, roughly 0.74 for fMRI and a good questionnaire.
• Median statistical power in neuroscience was estimated around 21 percent, which lowers the probability that a significant result is real and inflates published effect sizes.
• Seventy teams analysing one dataset split on the binary conclusion for five of nine hypotheses, so pipeline choice is itself a source of results.
• Brain-wide association effects are very small, and replicable estimates require thousands of participants rather than dozens.
• Within-subject effects with large true sizes, including retinotopy, face selectivity and the N400, replicate routinely; the crisis is concentrated elsewhere.
• The response has been preregistration and registered reports, permutation inference, data and map sharing, reporting standards, consortium datasets, and out-of-sample prediction.

Sources

1. Vul, E., Harris, C., Winkielman, P., & Pashler, H. (2009). Puzzlingly high correlations in fMRI studies of emotion, personality, and social cognition. Perspectives on Psychological Science, 4(3), 274-290. pubmed.ncbi.nlm.nih.gov
2. Button, K. S., Ioannidis, J. P., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S., & Munafo, M. R. (2013). Power failure: Why small sample size undermines the reliability of neuroscience. Nature Reviews Neuroscience, 14(5), 365-376. pubmed.ncbi.nlm.nih.gov
3. Botvinik-Nezer, R., Holzmeister, F., Camerer, C. F., Dreber, A., Huber, J., Johannesson, M., et al. (2020). Variability in the analysis of a single neuroimaging dataset by many teams. Nature, 582(7810), 84-88. pubmed.ncbi.nlm.nih.gov
4. Marek, S., Tervo-Clemmens, B., Calabro, F. J., Montez, D. F., Kay, B. P., Hatoum, A. S., et al. (2022). Reproducible brain-wide association studies require thousands of individuals. Nature, 603(7902), 654-660. pubmed.ncbi.nlm.nih.gov
5. Poldrack, R. A., Baker, C. I., Durnez, J., Gorgolewski, K. J., Matthews, P. M., Munafo, M. R., et al. (2017). Scanning the horizon: Towards transparent and reproducible neuroimaging research. Nature Reviews Neuroscience, 18(2), 115-126. pubmed.ncbi.nlm.nih.gov
6. Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349(6251), aac4716. pubmed.ncbi.nlm.nih.gov

The claim

On 30 September 2011 the New York Times published an opinion piece by the marketing consultant Martin Lindstrom under the title You Love Your iPhone. Literally. It reported a brain-imaging study of sixteen participants who were shown images of an iPhone and played the sound of a ringing iPhone while being scanned. According to the piece, the participants' insular cortex became active, and since the insula is associated with feelings of love and compassion, the conclusion followed: people are not addicted to their phones, they love them.

Four days later the newspaper printed a letter signed by forty-five neuroscientists saying, in effect, that the reasoning does not work. This lesson is that letter, worked out in full, because the error it names is the most common one in the public presentation of functional imaging and it is not confined to newspapers.

Step 1: name the direction of the inference

An imaging experiment measures activation given a task. That is forward inference: the probability of activation given a cognitive process, written P(A given C). It is what the experiment was designed to estimate, and it is estimated from data the experimenter controlled.

The claim in the op-ed runs the other way. It observes activation and concludes which process was occurring: the probability of a cognitive process given activation, P(C given A). That is reverse inference, and it is not what the experiment measured. Russell Poldrack named and formalised the problem in 2006.

The point: The two probabilities are not interchangeable, and the gap between them is governed by how often that region activates when the process in question is absent.

Step 2: write down the arithmetic

By Bayes' rule, the probability that the process was present given the activation is:

P(C given A) = P(A given C) x P(C) / [ P(A given C) x P(C) + P(A given not-C) x P(not-C) ]

Three quantities are needed, and the op-ed supplies only an intuition about the first. P(A given C), how reliably the region activates when the process occurs. P(C), the prior probability that the process was occurring at all. And P(A given not-C), how often the region activates when the process is absent, which is the term that does all the damage.

Work an illustrative case. Suppose the region activates in 80 percent of studies where the process of interest is present, the prior for the process is 0.5, and the region also activates in 60 percent of studies where the process is absent, because it is engaged by many things. Then:

P(C given A) = (0.8 x 0.5) / (0.8 x 0.5 + 0.6 x 0.5) = 0.40 / 0.70 = 0.57

The prior was 0.5 and the posterior is 0.57. Observing the activation moved the belief by seven percentage points. That is the whole evidential value of the observation, and no amount of confident prose changes it. Now suppose instead the region activates in only 5 percent of studies where the process is absent:

P(C given A) = (0.8 x 0.5) / (0.8 x 0.5 + 0.05 x 0.5) = 0.40 / 0.425 = 0.94

Same forward probability, same prior, completely different conclusion. Selectivity, not responsiveness, is what licenses a reverse inference.

Step 3: get the real base rate

The illustrative numbers above are not needed, because the field has a database for exactly this. Tal Yarkoni, Poldrack and colleagues built Neurosynth by automatically extracting activation coordinates and term frequencies from thousands of published fMRI papers, which makes both directions of inference computable.

Look up the insula there and the problem is immediate: it is among the most frequently reported structures in the entire imaging literature, appearing in a large fraction of published studies across pain, disgust, interoception, taste, empathy, risk, effort, uncertainty, salience, and motor tasks. P(A given not-love) is very high, which is the first case in step 2, not the second.

Neurosynth makes this concrete with two maps for every term. The uniformity map shows where studies using the term tend to report activation, which is forward inference. The association map shows where activation is reported more consistently for studies using the term than for studies not using it, which is reverse inference with the base rate divided out. For many terms the two maps look very different, and the second is much sparser. That difference is the lesson.

Step 4: rewrite the claim so it is true

Here is what the reported data can support, stated in a way that would survive review:

Images and sounds of a familiar mobile phone elicited a response in insular cortex, a region engaged by a very wide range of tasks including interoception, salience detection, effort and pain. The data do not distinguish affection from irritation, craving, or simple heightened attention, and no comparison condition was reported that would allow such a distinction.

Notice what the rewrite required. It named the base rate problem, it named the missing control, and it declined to name an emotion. It is duller, and it is what happened.

Step 5: when reverse inference is legitimate

The point is not that reverse inference is forbidden. It is a normal and useful form of reasoning when its conditions are met.

1. The region is genuinely selective. Primary visual cortex activation is decent evidence of visual stimulation. The fusiform face area is reasonably selective for faces, which is why the argument in lesson 7 is about degree rather than existence.
2. Selectivity is quantified, not asserted. A meta-analytic database gives you the posterior rather than an impression, and it is a five-minute check.
3. The inference is multivariate. Decoding from a distributed pattern, validated on held-out data, is a far stronger form of the same reasoning: it asks whether the pattern discriminates the states in question rather than whether one region lit up. Poldrack's 2011 follow-up frames the whole problem as one of large-scale decoding rather than of single regions.
4. The candidate set is constrained. Reverse inference is much stronger when the question is which of two specified processes occurred than when it is which of all possible mental states occurred.

Step 6: four more claims to debug

Worth holding on to: Every one of these claims is made in the peer-reviewed literature, not only in newspapers. The test is always the same: how often does this region activate when the proposed process is absent?

A second worked case, briefly

In November 2007 the same newspaper published a piece reporting brain scans of swing voters, in which amygdala responses to candidate images were read as anxiety and insula responses as disgust. A letter from seventeen neuroscientists followed, making the identical point. The recurrence is the interesting part: the error is not a one-off lapse but a standing feature of how imaging results get translated, because the vocabulary of a region invites a claim about a mental state, and the arithmetic that would check the claim is invisible in the picture.

Common misconceptions

• Reverse inference is always invalid. It is valid in proportion to the selectivity of the region, and the selectivity is measurable.
• A strong activation licenses a stronger inference. Magnitude is irrelevant to the Bayesian calculation; only the base rate of activation without the process matters.
• Naming a region is a finding. The region's name carries no information about its computational role, and many region labels are historical accidents.
• Multivariate decoding removes the problem. It strengthens the inference considerably and still requires an independent test set and a constrained set of candidate states.
• The error is confined to journalism. Peer-reviewed papers make the same move routinely, usually in a discussion section, and usually without any base rate.

Putting it together

• Forward inference estimates activation given a process; reverse inference claims a process given activation, and only the first is what an experiment measures.
• Bayes' rule shows the posterior depends on the probability of activation when the process is absent, which is the quantity these claims never report.
• With a region that activates in most studies regardless of process, observing its activation moves belief by a few percentage points at most.
• Neurosynth computes both directions from thousands of papers, and its association maps are reverse inference with the base rate divided out.
• The insula is among the most frequently reported structures in the literature, which makes any inference from insula activation to a specific feeling nearly worthless.
• Reverse inference becomes legitimate with quantified selectivity, multivariate decoding validated out of sample, and a constrained candidate set.
• Amygdala to fear, cingulate to conflict, dopamine to pleasure, and temporoparietal junction to mentalising are the four standing examples of the same error in the primary literature.

Sources

1. Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences, 10(2), 59-63. pubmed.ncbi.nlm.nih.gov
2. Poldrack, R. A. (2011). Inferring mental states from neuroimaging data: From reverse inference to large-scale decoding. Neuron, 72(5), 692-697. pubmed.ncbi.nlm.nih.gov
3. Yarkoni, T., Poldrack, R. A., Nichols, T. E., Van Essen, D. C., & Wager, T. D. (2011). Large-scale automated synthesis of human functional neuroimaging data. Nature Methods, 8(8), 665-670. pubmed.ncbi.nlm.nih.gov
4. Neurosynth. (n.d.). Meta-analytic term maps: insula. neurosynth.org
5. Lindstrom, M. (2011, September 30). You love your iPhone. Literally. The New York Times, opinion section. (The claim analysed in this lesson; answered in the same paper on 4 October 2011 by a letter signed by forty-five neuroscientists.)