Attention and the Brain

The bandwidth problem attention exists to solve

Begin not with a definition but with an arithmetic embarrassment. The two retinas together hold well over a hundred million photoreceptors, and the optic nerves carry roughly a million fibres from each eye towards the brain. The cochlea is resolving sound across the audible spectrum. The skin is reporting pressure, temperature and the position of every limb. All of it arrives continuously, in parallel, and none of it stops when you are busy.

Now consider what the brain can do with information. Recognising an object, understanding a sentence, deciding on an action, holding a plan in mind: these are expensive operations, and the systems that perform them are strikingly limited. Working memory holds only a handful of items. You cannot read two sentences at once, or fully identify two spoken messages arriving together.

Put the two facts side by side and a design problem appears. The input is enormous; the capacity is small. Something must stand between them and decide what gets through. That something is attention.

State the consequence plainly at the outset, because everything else on this page follows from it. If attention is a selection mechanism operating under a hard capacity limit, then what it does not select is not merely noticed less vividly. It may not be perceived at all. Seeing is not the passive arrival of light on a retina followed by automatic awareness; seeing is what happens to the fraction of the input that attention decided to spend processing on. The rest can walk through the middle of the scene and leave no trace, as the gorilla experiment below demonstrates.

William James put the psychological side of this in 1890, and the phrasing has never been bettered: attention, he wrote, is the taking possession by the mind of one out of several simultaneously possible objects or trains of thought, and it implies withdrawal from some things in order to deal effectively with others. Note the second clause. The withdrawal is not a side effect. It is the point.

Four kinds of selection, four different jobs

Attention is not one capacity but a family of them, and the members are separated by the problem each one solves. Grouping them by function, rather than by the tasks psychologists happen to use, keeps the distinctions honest.

Two of these need a sentence of mechanism, because a description is not an explanation. Vigilance decays, reliably and steeply: performance on a monitoring task falls within the first half hour, an effect called the vigilance decrement, and it is worst when the target is rare and the task monotonous. The best account is not that attention runs out like a battery, but that sustaining top-down selection against an unrewarding, unchanging input is itself effortful, and arousal drifts downward when nothing happens to lift it. The corollary is uncomfortable: the rarer the event you are watching for, the more likely you are to miss it, which is precisely when missing it matters most.

Divided attention is not one thing either. Two tasks interfere most when they compete for the same resources and least when they do not. Listening to speech while reading text interferes badly, because both draw on language processing; listening while walking barely interferes at all. Interference is the fingerprint of a shared bottleneck, and by mapping which pairs of tasks interfere, experimenters map where the bottlenecks are.

The cocktail party: how the ignored ear gave attention away

The modern science of attention begins with an engineering question. In 1953, Colin Cherry, at Imperial College London, posed what he called the cocktail party problem: how does a listener follow one conversation in a room where several people are talking at once, when a microphone in the same room records only an unusable mush? The question was practical. The method he devised to attack it became the most productive paradigm in the field.

Cherry used dichotic listening. He fed one spoken message into a listener's left ear and a different message into the right, then asked the listener to shadow one of them: to repeat it aloud, word for word, as it arrived. Shadowing is demanding enough to guarantee that attention really is on the assigned channel, and that is the point of it. The experimenter now controls where attention is, and can ask what became of the channel that was ignored.

The answer is the finding on which everything after it is built. Listeners who had shadowed one ear for minutes could report almost nothing about the other. They could not say what language the ignored message was in, and could not report a single word of its content. They had not heard it in any sense that matters.

But they were not deaf to it. They reported the physical properties of the ignored channel easily: whether the voice was male or female, whether it was speech at all or a pure tone, whether it changed from one to the other. The pattern was sharp and repeatable. Physical characteristics of the unattended channel survive. Meaning does not.

That contrast is the entire evidence base for the theories that followed. Whatever attention is doing, it draws a line, and the line falls somewhere between the crude physical registration of a sound and the extraction of what the sound means.

Early filter, and the finding that broke it

Donald Broadbent, at the Applied Psychology Unit in Cambridge, took Cherry's result and gave it a mechanism. His proposal, from the late 1950s, is the early selection or filter model, and its logic is worth reconstructing because it shows how a good theory works and how it dies.

The argument runs as follows. Incoming sensory information is registered briefly in a sensory buffer. Beyond the buffer sits a channel of limited capacity, the one that does the expensive work of identifying meaning. Because that channel cannot take everything, a filter stands in front of it. The filter selects on simple physical properties, which is all that is available before meaning has been extracted: which ear the sound arrived at, its pitch, the voice that spoke it. Whatever passes the filter is analysed for meaning. Whatever does not is blocked, and never analysed at all.

This explains Cherry's data exactly. Physical properties of the ignored channel are known because the filter itself operates on them, so they must be registered. Meaning is unknown because the channel was blocked before the meaning-extracting stage was reached. The model is elegant, testable, and makes a strong prediction: nothing from the unattended channel should ever be understood, because it never reaches the machinery that understands.

That prediction is false, and the way it failed is instructive.

Notice what this does and does not overturn. It does not restore the idea that everything is fully processed: listeners still cannot report the content of the ignored message. It shows that the filter is not a wall. Something gets through, and what gets through is exactly the material important enough to be worth an exception.

Anne Treisman supplied the repair, and it is one of the cleanest theoretical corrections in cognitive psychology. In her attenuation model, the filter does not switch the unattended channel off. It turns it down. The ignored message still travels onward towards the systems that recognise words and meanings, but weakened. Whether it then registers depends on a second factor: how easily each word is triggered. Treisman proposed that stored words have thresholds, and that the thresholds are not equal. Words that are important, or expected in context, have permanently low ones. Your own name has the lowest of all.

Now everything fits. The ordinary content of the ignored message is attenuated below the thresholds of the ordinary words it contains, so it is never recognised and nothing of it can be reported. Your name is attenuated by exactly the same amount, but its threshold is so low that even the weakened signal clears it, so it breaks through. One mechanism produces both the near-total ignorance and the striking exception, which is the mark of a theory that has explained something rather than relabelled it.

A third position, late selection, held that everything is fully processed for meaning and that attention selects only what reaches awareness and response. The modern verdict falls between the poles, and is best stated as a rule rather than a fixed point: selection happens early when the task loads perception heavily, and later when perceptual load is light and spare capacity spills onto the ignored material. The filter's position is not fixed by anatomy. It moves with the demand.

Two networks, because one would not survive

Cognitive psychology established that selection happens and roughly where in the processing stream it acts. Functional imaging, from the 1990s onward, established which brain systems do it. The organising framework, and the one that has held up, was set out by Maurizio Corbetta and Gordon Shulman in a 2002 review in Nature Reviews Neuroscience. Their claim was that attention is not implemented by one system but by two, and that the two do different jobs.

The division is not a taxonomic nicety. It is forced by a problem that neither system can solve alone, and the argument is worth making in full.

Imagine an attention system that was purely top-down. It would be superb at what it was told to do. Given a goal, it would apply that goal relentlessly, filtering the world for the target and suppressing everything else. And it would be lethal. An organism that only ever saw what it was already looking for would never notice the thing it had no reason to expect: the movement in the grass, the car that was not there a second ago, the smell of smoke. It would search its way efficiently into a disaster.

Now imagine the reverse, a system that was purely bottom-up, driven entirely by whatever in the environment happened to be salient. It would never miss the sudden and the loud. It would also never finish anything. Every flicker at the edge of vision, every noise in the corridor, would capture it, and it could not hold a goal for the seconds required to complete a thought.

Either system alone is fatal in a different way. The first is blind to the unanticipated; the second is incapable of persistence. Survival requires both, and it requires them to be in tension: a top-down system strong enough to hold a goal against distraction, and a bottom-up system with enough authority to override the goal when the world produces something that genuinely warrants it. That tension is the architecture. The dorsal network holds the line; the ventral network is licensed to break it.

Two details make the account more than a story. First, the ventral network is not a general novelty detector: it does not respond to just any unexpected event, but preferentially to unexpected events that are behaviourally relevant. This matters, because a circuit-breaker that fired for every irrelevant surprise would reintroduce the problem it was meant to solve. Second, the two networks are connected, and the dorsal system appears able to suppress the ventral one during demanding search, which is the mechanism of concentration: you become harder to interrupt not because the interruptions stop arriving but because the circuit-breaker has been turned down.

An older and complementary framework, developed by Michael Posner and Steven Petersen and revisited by them in 2012, carves attention into three functions rather than two networks: alerting (achieving and maintaining a state of readiness), orienting (selecting information from sensory input), and executive control (resolving conflict among responses). The two schemes are not rivals. Orienting maps onto the dorsal and ventral networks, executive control onto the frontoparietal machinery described under executive function, and alerting onto the arousal chemistry we turn to next.

Below the cortex: the gate and the chemistry

Cortical networks decide what should be selected. They do not do the selecting on their own. Two subcortical contributions are indispensable, and an account that omits them cannot actually work.

The thalamic gate. Almost all sensory information travelling to the cortex passes through the thalamus first, and the thalamus does not merely pass it on. A relay nucleus receives only a minority of its synapses from the sense organ it serves; most of its input arrives from the cortex, the brainstem, and the thalamic reticular nucleus, a thin inhibitory shell of GABA-releasing neurons wrapped around the thalamus and pointing inward at it. That shell is a control surface: it can damp the flow through one nucleus while leaving another open, which is exactly the operation selective attention requires. The top-down bias generated in parietal and frontal cortex therefore has a physical mechanism through which to act on sensory transmission itself, before the signal has reached the cortex at all.

The pulvinar, the largest nucleus of the human thalamus, is the second player. It is reciprocally connected with the visual, parietal and temporal association cortices and is implicated in visual attention and in coordinating activity between the areas it connects. The plausible reading is that it helps regulate which cortical areas are talking to each other, and how strongly, which is what a system that must prioritise one representation over another needs.

The chemistry of readiness and selectivity. A gate is useless if the system behind it is not in a state to act. That state is set by neuromodulators, chemicals released diffusely across wide areas of cortex, which change how the receiving circuits behave rather than carrying a message of their own.

Noradrenaline, released from the locus coeruleus in the brainstem, is the arousal and alerting signal. Its projections reach almost the whole cortex, and its activity tracks vigilance: too little and the system is drowsy and misses signals, too much and it is jumpy and distractible, with best performance in between. This inverted-U relationship between arousal and performance is one of the oldest findings in psychology and one of the most useful, because it explains why exhaustion and panic both degrade attention by different routes. A warning cue that says something is about to happen works, in part, by driving this system.

Acetylcholine, released from the basal forebrain, does something more specific: it enhances the signal-to-noise ratio in sensory cortex, sharpening the response to the attended stimulus relative to the background. Where noradrenaline sets how alert you are, acetylcholine helps set how sharply you can distinguish what you are attending to from what you are not. Alerting and selecting are chemically dissociable, which is one reason the psychological distinction between them is more than a convenience.

The gorilla argument: failures are the proof

If attention really is selection under a hard capacity limit, there ought to be a dramatic and reproducible consequence: people should fail to see things that are fully visible, directly in front of them, and entirely unmissable, provided their attention is engaged elsewhere. The prediction sounds implausible. It is correct.

In 1999, Daniel Simons and Christopher Chabris published a study in Perception titled Gorillas in our midst. Observers watched a short video of two teams, one in white shirts and one in black, passing basketballs among themselves, and were told to count the passes made by one team, a task demanding enough to occupy attention thoroughly. Partway through, a person in a full gorilla suit walked into the middle of the scene, stopped, faced the camera, thumped their chest, and walked off. The gorilla was on screen for several seconds, unoccluded, in the centre of the display.

Roughly half of the observers did not see it. Asked afterwards whether they had noticed anything unusual, they said no. Shown the video again without a counting task, they saw the gorilla at once, and frequently refused to believe it was the same video.

Be precise about what this proves, because the result is often quoted loosely. It does not show that people are careless, or that the video was tricky. The gorilla's image fell on the retina and the signal travelled the ordinary pathway. What did not happen was selection, and without selection there was no perception. Inattentional blindness is the name for it: the failure to see a fully visible, unexpected object because attention was occupied elsewhere.

The companion phenomenon is change blindness, which makes the same point from another angle. If a large element of a scene is altered during a brief interruption, a flicker, an eye movement, a cut in a film, observers routinely fail to notice, even when they are looking for changes and even when the change is enormous. In one well-known demonstration, an experimenter stopping a passer-by for directions was swapped for a different person behind a passing obstruction, and a substantial proportion of those passers-by carried on the conversation without registering that they were now talking to a stranger.

Both converge on the conclusion the whole page has been building towards. Vision is not a recording. You do not hold a detailed internal copy of the scene against which changes could be checked. What you hold is the product of attention: a rich representation of the few things you selected, and, for everything else, gist and assumption. The feeling of seeing everything is genuine, and it is an illusion. Attention makes it possible to see anything at all, and its price is everything you were not attending to.

Neglect: when half the world stops mattering

The laboratory demonstrations show attention failing in healthy people. The clinic shows what happens when the attention machinery itself is damaged, and the resulting syndrome is the strongest single argument that attention is a real, separable system rather than a name for perception working well.

Hemispatial neglect most often follows damage to the right parietal cortex or the right temporoparietal junction, typically from a stroke, which is precisely the ventral network described above. The patient fails to attend to the left side of space, and the presentation is startling. Asked to copy a clock, they draw the circle and crowd all twelve numbers into the right half. Asked to cross out every line on a page scattered with lines, they cross out only those on the right. Given a plate of food, they eat the right half and report that they have finished. Some shave only the right side of the face, or dress only the right side of the body.

Three features carry the theoretical weight, and each rules out an alternative explanation.

Set that last point against the gorilla study and the two illuminate each other. In both cases the information is physically available, in both cases it is not experienced, and in both cases the person is confident they saw what there was to see. Neglect is inattentional blindness made permanent and spatial by a lesion, and the same signature appearing in a damaged brain and in a healthy brain under load is strong evidence for one mechanism: selection, doing its job, and revealing itself only in what it leaves out.

The right-lateralisation also explains a clinical asymmetry that would otherwise look arbitrary. Severe, persistent neglect follows right-hemisphere damage far more often than left. If the right hemisphere's ventral network attends to both sides of space while the left hemisphere's attends chiefly to the right, then left-hemisphere damage leaves the right side of space still covered by the right hemisphere, while right-hemisphere damage leaves the left side covered by nothing.

Multitasking is switching, and switching is taxed

Everything above has one practical consequence, and it is the one most often denied.

The honest formulation, then, is this: multitasking is task-switching with a hidden tax. The switching is real, the tax is real and quantifiable, and the hiddenness is not incidental. It is guaranteed by the architecture. The single most useful thing this page can offer a reader is that sentence, and its corollary, which is that the way to protect attention is not to try harder while divided but to reduce the number of things competing for selection in the first place.

Attention in the wild

Every mechanism on this page shows up in ordinary experience, and recognising them is the test of whether the account has been understood.

Related topics