Reward, error and self-play: what machine learning shows about human learning
October 5, 2026 · 16 min read
Michael Littman's history of reinforcement learning, Terry Sejnowski's account of the brain's reward and value signals and Fei-Fei Li on how AI learns and where minds differ describe learning in machines and people in overlapping terms. The comparison clarifies which lessons from AI transfer to human learners and which do not.
The rules, numbers and steps worth keeping in front of you.
- Sejnowski: the brain learns by comparing the reward received with the reward expected, then adjusting.
- He says the same rule runs in flies, in people and in AlphaGo.
- Sejnowski: punishment produces one-trial learning; that is a warning about trauma, not a teaching method.
- Learning has to be active; being told the answer stores a fact that will not transfer.
- Pair instruction with practice: lectures plus problem sets.
- Self-test to locate errors, not to grade yourself.
- Littman: systems learn basic play fast because they are told who won, so seek clear feedback.
- Littman: self-play improves one game and nothing else, and returns diminish.
- Littman: learners need to be pushed back at; argue and lose a few times.
- Li: a child learns a cat from three to ten examples; a model needs vast data.
- Li: a machine's objective is not a feeling; a person has to want to learn.
- Sejnowski: doctors using AI reached 98 percent accuracy, against about 90 percent for either alone.
- Think about a hard problem before sleep; sleep consolidates the day's experience.
- Decisions about sleep medication belong with a doctor.
This is the short form. The full insight has every expert's reasoning, their own words and the sources.