Dr. Terry Sejnowski: How to Improve at Learning Using Neuroscience & AI
Learning is often treated as a matter of absorbing more information. In his conversation with Andrew Huberman, computational neuroscientist Dr. Terry Sejnowski offers a more useful model: lasting knowledge develops through prediction, action, feedback, and biological consolidation. Reading and listening matter, but they are only part of the process.
Sejnowski draws connections among neuroscience, artificial intelligence, education, sleep, exercise, and motivation. His central message is encouraging: the brain remains adaptable throughout adulthood, and learning can improve when we work with its systems instead of relying on passive exposure or willpower alone.
Understanding the Brain at the Algorithmic Level
Sejnowski directs the Computational Neurobiology Laboratory at the Salk Institute, where his work combines mathematics, neuroscience, computing, and artificial intelligence. That perspective shapes how he approaches one of science’s hardest questions: how does the activity of cells and circuits become perception, thought, and behavior?
A detailed inventory of neurons, molecules, and connections does not by itself explain how the brain operates. Conversely, observing behavior from the outside cannot fully reveal how the underlying machinery produces it. Sejnowski describes an intermediate, algorithmic level between these two views.
An algorithm is comparable to a recipe: it specifies the operations a system performs to reach an outcome. In neuroscience, identifying such operations can explain how circuits transform experience into better decisions or more skilled behavior. This level is also where neuroscience and AI increasingly overlap. Both fields ask how a system can learn from feedback, generalize beyond previous examples, and select actions that move it toward a goal.
The Basal Ganglia, Dopamine, and Learning from Outcomes
One important learning system involves the basal ganglia, a collection of brain structures that helps refine sequences of actions. Practice allows these circuits to make movements smoother and more automatic, whether someone is improving a tennis stroke, a golf swing, or another complex skill.
Sejnowski argues that the same broad machinery also contributes to cognitive expertise. Becoming fluent in medicine, finance, physics, or neuroscience requires learning productive sequences of thought—not merely storing isolated facts.
The basal ganglia help organize both “go” and “no-go” signals. Effective performance depends on initiating the right action while suppressing competing actions. A surgeon, for example, must execute a precise procedure while excluding irrelevant thoughts. Huberman and Sejnowski extend this distinction to adolescence, when systems involved in restraint and inhibition are still developing.
Dopamine plays a central role in updating these circuits. The brain predicts what reward should follow an action, compares that forecast with what actually happens, and adjusts its connections when the result differs from expectation. Over repeated experiences, this process builds what computational researchers call a value function: an accumulated estimate of which choices are likely to produce beneficial or harmful outcomes.
This principle is not confined to the laboratory. It influences habits, professional judgment, social behavior, and personal preferences. Small positive results can gradually reinforce a useful pattern, while a strongly negative event may produce rapid learning after only one exposure.
Related reinforcement-learning methods helped make systems such as AlphaGo possible. The comparison does not mean that brains and AI are identical. Rather, it shows how a relatively compact learning rule can produce sophisticated behavior when it operates repeatedly across many situations.
Knowing Something Is Not the Same as Being Able to Do It
A major theme of the discussion is the difference between cognitive and procedural learning. Cognitive learning provides explicit knowledge: concepts, explanations, rules, and facts. Procedural learning turns that knowledge into an ability that can be used efficiently.
A person cannot learn tennis solely by reading about technique. Nor can a pilot become competent through textbooks without practical training. Even academic subjects require procedural development. Memorizing an equation is different from recognizing when to use it, applying it to an unfamiliar problem, and checking whether the answer makes sense.
The two systems should therefore support each other. Understanding why a method works can guide practice, while practice exposes gaps that remain hidden during passive study. Sejnowski is critical of educational approaches that minimize homework, problem-solving, or other forms of active rehearsal. Without repeated application, students may recognize material without developing command of it.
A productive learning cycle looks like this:
Study the underlying idea.
Attempt a problem without immediately consulting the answer.
Identify the mismatch between the prediction and the result.
Correct the method and try again in a new context.
Return to the material after time has passed.
Testing is valuable within this cycle because retrieval is itself a learning event. It reveals errors, strengthens access to knowledge, and trains the learner to produce an answer rather than merely recognize one.
Active Learning Builds Generalization
Sejnowski emphasizes that the brain does not store information like a computer copying files. Learning requires interaction: attempting, predicting, making mistakes, and revising behavior.
This active process is what supports generalization—the ability to use prior learning in a situation that is not exactly the same as the training example. It is also a defining goal of modern AI. A language model that only reproduced its training data would be far less useful than one capable of responding coherently to a novel question.
Humans can sometimes generalize from remarkably few examples, although that ability depends on prior experience and existing mental structure. The practical implication is that learners should vary their practice. Solving the same type of problem repeatedly may improve familiarity, but mixing contexts and confronting unfamiliar cases provides a better test of transferable understanding.
Sejnowski and educator Barbara Oakley developed the free online course Learning How to Learn around many of these principles. Although initially conceived with students in mind, it attracted a broad international audience, including working adults who needed to acquire new skills. Its popularity reflects a modern reality: formal education ends, but the need to learn does not.
Memory Is Distributed Across the Brain
Older accounts of brain function often assigned each task to a narrowly defined region: visual information belonged to visual cortex, auditory information to auditory cortex, and movement to motor areas. Those distinctions remain useful, but newer recording methods show far more widespread coordination.
Sejnowski discusses research in which activity related to a task appears across multiple regions, sometimes with substantial input from areas that would not traditionally be considered primary for that task. Complex cognition seems to emerge from interactions among distributed systems rather than from a single storage location.
Recordings from people with epilepsy have offered an especially valuable window into these dynamics. Some patients receive implanted electrodes as part of clinical care, creating rare opportunities for researchers to observe human neural activity directly. With appropriate research collaboration, scientists can study how activity changes while participants learn and while they sleep.
These findings make memory look less like an object placed in one compartment and more like a pattern integrated into an existing network.
Sleep Consolidates What Practice Begins
Learning does not stop when a study session ends. During non-REM sleep, the brain produces brief bursts of coordinated activity known as sleep spindles. Sejnowski describes these events as part of the process through which recent experience becomes incorporated into longer-term cortical knowledge.
The hippocampus plays a temporary organizing role, replaying elements of recent experience during sleep. This replay helps integrate new information without simply overwriting what was already learned. Adequate sleep is therefore not just recovery after studying; it is an active stage of memory formation.
The conversation also examines research involving zolpidem, commonly known as Ambien. In an experiment discussed by Sejnowski, the drug increased sleep spindles and improved retention of previously learned material. Yet zolpidem can also impair memory for events occurring after it is taken.
That tradeoff illustrates a broader warning: manipulating one biological mechanism may produce benefits and costs elsewhere. Sejnowski applies the same caution to nootropics and hormone-based interventions. Pharmacological enhancement should not be mistaken for a free improvement to an isolated system, because the brain and body maintain interconnected forms of balance.
Exercise Supports Learning and Creative Thought
Sejnowski uses physical activity as a cognitive tool. Exercise can prepare the brain for learning, influence sleep physiology, and create conditions in which ideas combine more freely.
The discussion notes an association between daytime exercise and sleep spindles later that night. Exercise also challenges motor systems, while sleep helps refine and consolidate their adaptations. This relationship reinforces the idea that cognitive performance cannot be separated neatly from physical behavior.
For learners, the lesson is straightforward: movement is not necessarily time taken away from intellectual work. Regular exercise may support the processes that make study productive, particularly when paired with sufficient sleep.
Connectivity, Psychedelics, and Scientific Caution
Sejnowski also discusses psilocybin research, including clinical studies reporting improvement in some people with major depression. Brain imaging suggests that psilocybin can temporarily increase communication among regions that are usually more segregated, producing a pattern with certain similarities to the broad connectivity seen during REM sleep.
This raises intriguing questions. Development involves both forming connections and pruning them, partly because maintaining neural connections is metabolically expensive. Learning a new skill may require new coordination among previously separate systems, but more connectivity is not automatically better.
The discussion presents this area as an evolving scientific problem, not a do-it-yourself learning protocol. Any possible therapeutic benefit must be considered alongside clinical supervision, individual risk, legal constraints, and the limits of current evidence.
A Practical Neuroscience-Based Learning Routine
The most useful ideas from the conversation can be combined without turning learning into an elaborate optimization project. Begin with focused instruction, but move quickly into active use. Attempt retrieval before reviewing, solve problems that vary in form, and treat errors as information for updating the next attempt.
Keep rewards small and immediate enough to reinforce progress. Practice regularly so that effortful steps can become procedural. Use exercise to support cognition and protect enough sleep for consolidation. Above all, judge learning by what you can generate, explain, or perform in a new context—not by how familiar the source material feels.
Sejnowski’s account replaces a simplistic “study harder” message with a biological cycle: predict, act, receive feedback, update, rest, and return. Learning becomes more effective when knowledge, behavior, and recovery are treated as parts of the same system.
