How to Improve at Learning Using Neuroscience & AI | Dr. Terry Sejnowski

The basal ganglia implement a simple reinforcement learning rule: predict future reward, compare it to actual reward, and update synapses based on the error signal. Over time this builds a ‘value function’ that encodes what is good or bad for you—across motor skills (tennis serve), thinking (problem-solving), and social behavior (relationships, culture, language norms). Both positive rewards and punishments train this system; negative, high-salience events (e.g., trauma, bad relationships) can create powerful one-trial learning that shapes behavior for decades (including PTSD-like patterns). Practically, you can harness this by deliberately noticing what felt better or worse than expected and adjusting your actions and strategies iteratively.

The brain has two major learning systems: a cortical, cognitive system (explicit knowledge, rules, concepts) and a subcortical, procedural system (basal ganglia; automatic skills). Classroom lectures and reading mainly train the cognitive system; problem sets, drills, and real-world practice train the procedural system. Sejnowski and Oakley’s MOOC “Learning How to Learn” (free on Coursera) shows that active recall, problem-solving, and spaced practice are essential—merely hearing or rereading information does not build the robust, generalizable competence needed for physics, medicine, or even basic numeracy. Policies that remove practice because it is ‘stressful’ deprive students of the only mechanism that automates skill and makes cognition efficient.

During non‑REM sleep, brief circular traveling waves called sleep spindles (1–2 seconds long) help transfer recent experiences from the hippocampus into stable cortical memory without overwriting existing knowledge. More (and higher-quality) spindles correlate with better next‑day recall. Zolpidem (Ambien) roughly doubles spindle count and can double memory consolidation for material learned before taking it—but it impairs encoding of new experiences after ingestion, explaining blackout-like episodes in travelers. REM sleep, by contrast, appears especially important for motor system tuning and perhaps emotional processing. The practical implication: prioritize sufficient nightly sleep, avoid chronic REM/spindle suppression (e.g., via some drugs or substances), and use behavioral tools (exercise, consistent schedules) before relying on pharmacology.

Mitochondria provide cellular energy (ATP) and degrade with age and certain medications, contributing to reduced vigor and cognitive slowing. Regular physical exercise increases mitochondrial number and efficiency across organ systems, improves neurogenesis (e.g., in hippocampus), enhances synaptic plasticity, and supports sleep quality—all of which boost learning and cognitive resilience. Interval-style efforts (short all‑out sprints interspersed with easy movement) can be particularly potent for muscle and cardiovascular adaptations. Long-term education and continuous cognitive engagement likewise build a ‘cognitive reserve’ that delays clinical onset of dementia, suggesting that mental and physical ‘training’ throughout life are both forms of energy and resilience banking.

Modern AI systems (transformers, large language models) learn by predicting the next word and, in doing so, build internal semantic models of language and world structure. They implement algorithms akin to those in the brain’s circuits (e.g., reinforcement learning in AlphaGo mirroring basal ganglia; self‑attention mirroring temporal context operations the brain must perform). Doctors already achieve higher diagnostic accuracy when they use AI as an assistant (e.g., dermatologists plus AI reaching ~98% accuracy vs. ~90% alone). Researchers like Sejnowski’s colleague Rusty Gage use AI as an ‘idea pump’ to propose novel experiments from existing data. The near‑term opportunity is symbiosis: humans provide depth, judgment, and values; AI provides breadth, speed, and pattern discovery.