How We Learn: Why Brains Learn Better Than Any Machine . . . for Now

By Stanislas Dehaene

9 min read

Why does it matter? Because your assumptions about how children learn are built on a model of the brain that neuroscience has quietly demolished.

At seven, Felipe is blind, tetraplegic, and writing novels in three languages. That's not there to inspire you. It's there to make you ask: what is a brain actually doing, and how does one in those circumstances still arrive at language, at story, at three of them?

Learning was never optional, never supplementary: it was the only way biology could bridge the gap between what genes specify and what a brain becomes. Which changes the question entirely. It's not whether you can learn more. It's whether you understand the mechanism well enough to stop undermining it — and whether the ways we've been teaching children have been working with that mechanism or, quietly, against it.

Your Genome Is a 750-Megabyte CD-ROM — and That's Why Your Brain Had to Learn Everything Else

In September 2009, Dehaene walked down a hospital corridor in Brasília, steeling himself. He was about to meet Felipe, a seven-year-old who had been shot at age four, leaving him fully blind, nearly paralyzed from the neck down, breathing through a surgical opening in his throat. Three years in a hospital bed.

The door opened. Felipe was alive with questions, asking Dehaene mischievous things about French words. He spoke three languages fluently. He had dictated novels to an assistant, then learned to type them himself on a custom keyboard. The hospital team printed his stories as tactile books with embossed illustrations, which Felipe traced with whatever fingertip sensation remained.

The question that gripped Dehaene wasn't how Felipe had survived. It was how a brain stripped of sight, movement, and nearly every ordinary childhood input had still arrived at language, story, and curiosity. The standard picture of learning, that the brain starts empty and experience fills it, predicts a very different child.

That picture fails a simple arithmetic test. Your entire genome contains roughly three billion base pairs, encoding about 750 megabytes of information: the equivalent of a single old CD-ROM. Your brain holds roughly 100 trillion synaptic connections; at even one bit each, that's 100 terabytes, a hundred thousand times more information than the genome that built it. Genes cannot possibly pre-specify every synapse. What they can do is establish the architecture: the regions, the layers, the rough wiring plan. What fills those connections in has to come from somewhere else.

That somewhere else is learning. Learning wasn't supplementary — it was arithmetic necessity. The genome builds the palace; experience writes everything on the walls. Felipe's brain arrived with the structure intact — the capacity for language, the hunger for story, circuits ready to be shaped by input. What it needed was the world. It's how biology completes itself.

The Windows That Close: Why Two Weeks of Missing B1 Can Cost a Child Their Grammar

That learning, though, has a deadline.

The brain's plasticity is not a safety net. It is a series of windows, each open for a limited time, each governing something specific — and when they close, they close for good.

In early 2003, an Israeli formula manufacturer quietly removed thiamine (vitamin B1) from their soy-based infant formula to cut costs. The body doesn't store thiamine, so the absence hit fast. Within months, hundreds of babies arrived in hospitals with symptoms resembling adult alcohol poisoning: confusion, eye disorders, coma. Two died. Once doctors identified the cause and restored the vitamin, the acute crisis resolved in days. Between 600 and 1,000 infants had gone without thiamine for roughly two to three weeks during the earliest months of their lives.

Six years later, psychologist Naama Friedmann tested about sixty of them. Their IQs were normal. Vocabularies were normal. They could match a ball of wool to a sheep rather than a lion. But ask them to parse a sentence (to figure out who did what to whom) and most failed. Their syntactic grammar was severely impaired. A sentence like "Show me the girl the grandmother is combing" tripped them up: "the girl" is the object of the verb, not the subject, and working that out requires processing that had apparently depended on something missing in those first few months.

Two to three weeks. A couple of missing milligrams. A specific cognitive faculty, permanently compromised. Everything else intact.

The thiamine disaster illustrates the point precisely. It didn't impair a general learning ability — it closed a specific window at a specific time, and no amount of later language exposure reopened it. The same pattern holds across every domain Dehaene examines. Binocular vision must be calibrated before age three (children born with cataracts who have them removed too late rarely recover full depth perception). The phonological patterns of a native language lock in around twelve months, which is why adult Japanese speakers can spend decades in English-speaking countries without reliably distinguishing R from L. Grammatical learning stays open until around seventeen, then drops sharply. Early intervention isn't just more effective. For some faculties, it's the only kind that works at all.

Babies Aren't Empty Vessels — They Arrive Running a Bayesian Probability Engine

What do babies know before anyone teaches them anything? Quite a lot, it turns out.

Start with probability. Cognitive scientist Fei Xu showed eleven-month-olds a transparent box holding three red balls and one green one. A hand drew out a single ball. When it came out green — a one-in-four shot — babies stared significantly longer than when it came out red. Controls varying ball position and draw timing confirmed the pattern: the stare tracked improbability. Something in those small heads had already established expected frequencies and registered a deviation. Before they could walk or speak, they were running a Bayesian inference engine.

This probabilistic intuition belongs to a broader innate endowment: circuits for objects, quantities, physical causation, intentional agents. The number sense in particular has striking downstream consequences.

When neuroscientist Marie Amalric slid fifteen professional mathematicians into an MRI scanner and showed them formulas dense enough to unsettle most graduate students — things like ∫s∇×F•dS — the circuit that lit up was the same parietal network that activates when an infant tracks a small set of objects. Cultural learning at its most rarefied still runs on ancient circuitry.

Emmanuel Giroux, blind since age eleven, runs a sixty-person mathematics lab in Lyon. His brain shows the same parietal and frontal activation as sighted colleagues, and his visual cortex, freed from sight, gets recruited for math as well.

Dehaene calls this neuronal recycling: cultural skills colonize existing circuits rather than build new ones. Reading moves into regions originally organized for object recognition. Arithmetic extends circuits that evolved to track approximate quantities. Because that architecture is fixed — the same parietal network shows up in Giroux whether blind or sighted, in an infant or a Fields Medal winner — building new learning on top of existing intuitions is structural necessity. Your brain has no other way to absorb a genuinely new idea.

What Your Students Don't Attend To, They Cannot Learn — Even If Their Eyes Were Open

Imagine two groups of adults sitting down to learn an unfamiliar writing system: elegant, curving symbols, nothing like the alphabet they grew up with. Both groups study the same sixteen words. The only difference is what they're told to pay attention to.

The first group is instructed to treat each symbol like a Chinese character: one shape, one meaning, attend to the whole. The second group is told something different: these curves are actually three stacked letters, and they'll learn faster by attending to each one individually.

After training, psychologist Bruce McCandliss showed both groups sixteen new words in the same script and asked them to read aloud. The whole-word group read none of them — zero percent. The letter-focused group read 79 percent. Same exposure, same study time, radically different outcomes. Brain scans showed the two groups had activated entirely different circuits. The letter-attenders used the left-hemisphere ventral visual region (the brain's standard reading network). The whole-word group had activated the equivalent region in the right hemisphere: not a weaker version of the reading circuit, but the wrong circuit entirely.

That's what makes this more than a tip about classroom technique. Attention isn't a volume knob you turn up or down on learning. It's a routing switch. The brain contains multiple circuits capable of processing visual symbols, and attention determines which one gets the signal. Misdirect it, and you don't get poor learning in the correct place — you get competent learning in the neurologically wrong place, learning that cannot generalize because it's running on the wrong circuit.

The practical implication cuts deep. A student can sit in class, eyes forward, dutifully studying — and still learn nothing transferable, because their attention landed on the wrong feature of the material. What they cannot attend to correctly, they cannot learn correctly. The teacher's job, it turns out, isn't just to present information. It's to be a traffic controller for attention itself. Scale that misdirection up and the routing problem becomes a curiosity problem: a child who persistently attends to the wrong features of a subject eventually stops wanting to attend to it at all.

Curiosity Is a Computable Algorithm — and Exhaustive Teacher Demonstrations Break It

Developmental psychologist Laura Schulz gave kindergartners at MIT a contraption of plastic tubes hiding a mirror, a horn, a light toy, and a music box. Children who received it without comment dove in, rummaging and poking until they'd found nearly everything. A second group got a brief introduction: "Look, let me show you what it does," followed by a demonstration of the music box. Exploration collapsed. Shown one feature by a knowledgeable adult, the children largely stopped investigating the others — they inferred, reasonably enough, that an adult trying to be helpful would have demonstrated anything worth knowing. The demonstration had closed the inquiry.

Think of curiosity as a budget manager your brain runs continuously, allocating attention toward wherever the expected learning gain is highest. Frédéric Kaplan and Pierre-Yves Oudeyer, two French engineers, built a robot implementing exactly that: a world-prediction module, a second module tracking how fast those predictions were improving, and a reward circuit biasing the robot toward actions expected to maximize future learning. Set loose on a baby mat, it behaved exactly like an infant — spending minutes absorbed in a stuffed elephant ear, mastering everything it could extract, then losing interest and moving on. Familiar things bore: no gain expected. Completely alien things repel: too little traction to make progress. The sweet spot is intermediate complexity: challenging enough to yield something new, tractable enough to actually learn.

The teacher had been warm, clear, and completely counterproductive. The issue was signaled omniscience: the children's curiosity engines computed no remaining learning gain and powered down. The fix isn't to stop teaching. Leave deliberate gaps and the curiosity engine keeps running.

Zero Surprise Means Zero Learning — Why Errors Are the Signal, Not the Problem

Errors are not evidence of poor teaching or insufficient effort. They are the literal chemical signal the brain requires to change.

The neuroscientific proof is stark. In the 1970s, Rescorla and Wagner proposed that the brain learns only when its predictions are violated — and a single animal experiment makes this impossible to argue with. Train a dog to associate light with food. Once it's learned that, add a bell to the trials: light plus bell, then food, hundreds of times. Test the bell alone. The dog shows nothing: no salivation, no conditioned response, as if the bell-food pairing had never happened. The associationist view of learning, in which the brain passively records co-occurring events, predicts the opposite result. Hundreds of pairings should produce strong learning. But they produce zero, because the light already explains everything. The dog's brain generates no prediction error (it already knows what's coming), and without that signal, no synaptic update fires. The bell was present. The food was present. The pairing repeated hundreds of times. The association never formed.

Re-reading a chapter mimics that experiment exactly: everything is present, nothing is surprising, and your brain generates almost no update signal. Psychologist Henry Roediger showed that students who alternate study with self-testing (generating predictions, getting them wrong, receiving correction) substantially outperform students who only study, even when testing cuts into study time. The gap shows up on tests taken weeks later. Both students and teachers predict exactly the wrong outcome — that more time with the material means more learning. The illusion is that having information in front of you means you're learning it. What actually drives encoding is the gap between what you expected and what you got.

Errors are the brain's way of noticing it was wrong. That noticing is learning.

Sleep Is Not Recovery Time — It's When Your Brain Converts Experience Into Knowledge

What if the studying your student did last night was only half the job — and the other half happened while they slept?

In 1994, neuroscientists Wilson and McNaughton discovered that hippocampal neurons in sleeping rats fire in the same sequence they used while the rats explored a maze that day — but at twenty times the speed. The replay is accurate enough that they built a decoder and used it to read the content of rat dreams from brain activity alone. Every night, in accelerated bursts, the brain re-runs the day's experience.

Sleep researcher Jan Born measured what that replay produces. He taught volunteers a number-transformation algorithm containing a hidden shortcut — a trick that slashed the required steps. Before sleep, almost nobody found it. After one full night's sleep, twice as many discovered it spontaneously. Participants who stayed awake for an equivalent period found nothing. Elapsed time didn't generate the insight. Sleep did. The shortcut had always been in the material; overnight replay reorganized it into something the waking brain could suddenly see.

Dehaene explains why: the 20x acceleration compresses events separated by minutes or hours into an adjacent neural sequence, letting the cortex detect patterns invisible in real time. The hippocampus holds the raw episodes; repeated replay distills them into abstract cortical knowledge. What arrives in the morning is a compressed model, already beginning to generalize.

Which makes the school schedule a neuroscience problem. Puberty shifts adolescents' circadian clocks forward; they can't fall asleep early even when they try. Studies that delayed school start times by thirty to sixty minutes found more sleep, better attention, and higher grades. Zero cost, no new technology required. The redesign is already within reach.

Attend, Engage, Err, Sleep — The Algorithm You Were Born With

The algorithm is already running — it ran last night while you slept, and it ran every time you made a mistake today and felt the small sting of being wrong. Dehaene's thirteen recommendations collapse into four words: attend, engage, err, sleep. Not metaphors for good habits — the literal sequence of operations your brain executes every time it learns anything. What the science offers isn't a new system to install. It's a precise description of the one evolution already built, the same one Felipe was running from a hospital bed, dictating novels through a breathing tube. You don't need to redesign a school to start. Leave one deliberate gap in tomorrow's lesson. Correct an error within minutes, not weeks. Protect a teenager's eight hours tonight. You just have to stop working against it.