Reward, error and self-play: what machine learning shows about human learning
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.
Machine learning and human learning are described in similar words because part of the machinery really is shared. The brain's reward system and the program that beat the Go world champion use the same rule: compare the reward you got with the reward you expected, and adjust. What carries over to people is learning by attempting and being corrected, clear feedback, lots of practice alongside instruction, and opponents who push back. What does not carry over is the scale and the motive.
Terry Sejnowski, who directs a computational neurobiology laboratory at the Salk Institute, makes the case for sameness. The rule, he says, is "It's simply to predict the next reward you're going to get." Each action updates an accumulated sense of what is good or bad for you, carried by dopamine neurons. He says the same algorithm runs in flies, in people and in AlphaGo. He adds that punishment teaches in a single trial, which explains how strongly bad experiences mark us and is not a recommendation to teach harshly. He separates slow, step-by-step cognitive learning from automatic procedural learning, and says learning has to be active. Working problem sets, not memorizing equations, made him a physicist.
Michael Littman, a Brown University computer scientist, supplies the cautions. Early game-playing systems started from expert moves or human data before playing themselves, and the human skill behind those results is easy to discount. Self-play has a ceiling and a narrow reach: a system practicing against itself gets better at that game and nothing else. Speaking of an older language model, he said it captured the statistics of language, and that what a learner needs is "I think you really need to be pushed back at". Teaching his children to drive showed him how much of driving is social.
Fei-Fei Li, a Stanford computer scientist, stresses the difference. A child learns a cat from perhaps three to ten examples through a pathway nobody has fully explained, while a model needs enormous amounts of data. A machine's objective is not a feeling. A chatbot expresses sympathy from patterns in data, and a friend does so out of care.
They agree that learning is driven by the gap between prediction and outcome, that people and machines do best together, and that interaction with people matters. They disagree on whether the learning is the same or fundamentally different, on whether language models truly generalize, and on how impressive pure self-play is. Sejnowski and Li may both be right, since one describes a specific reward circuit and the other how today's large models are trained. The sources say less about a human version of self-play than about reward and error, and the article draws that parallel itself.
What to do with it: make an attempt or prediction before you see the answer, then compare. Pair every block of instruction with practice of your own. Test yourself to find errors, not to grade yourself. Arrange fast, clear feedback. Argue with people who disagree and expect to lose a few times. Vary what you practice against, because a narrow skill may not carry. Think about a hard problem before sleep, and walk away when stuck. Use AI as a partner or a tutor you question, not a substitute. Protect your motivation by limiting passive scrolling. Practice social skills with other people. Decisions about sleep medication belong with a doctor.