Korva is a scroll of the internet’s best ideas.
You learn at the edge of failing
Throwaway mistakes don't count: missing serves on purpose at a tennis court won't improve a serve. What works is serving as well as possible and failing anyway. The same holds for dance, a crossword, chess, coding, or any physical skill. That failure sends an error signal through circuits relayed by the cerebellum, the mini brain at the back, and the signal carries forward so the rest of the brain has to reshape itself.
The reshaping happens during deep rest, in both slow-wave deep sleep and REM sleep. Sleep needs vary, with some people needing six hours a night and others nine. A night or two of poor sleep doesn't wipe out learning, though doing that consistently takes a toll.
Why does the cerebellum care about errors?
The cerebellum compares the movement the brain intended with the movement that actually happened. Error signals travel on climbing fibers from the brainstem to Purkinje cells, the large output neurons of the cerebellum. In the classic Marr, Albus, and Ito theories, those signals weaken specific synapses so the next attempt comes out closer to the goal. The cerebellum holds roughly half of all the neurons in the human brain.
How does sleep lock in new skills?
In 1994, Matthew Wilson and Bruce McNaughton recorded hippocampal neurons in rats and published the results in Science. The same firing patterns that appeared while rats ran a maze reappeared during later slow-wave sleep. That replay is one proposed way the brain strengthens new memories overnight. Matthew Walker and colleagues also reported in Neuron in 2002 that finger-tapping speed in people improved after a night of sleep without any more practice.
Is there an ideal error rate for learning?
Robert Bjork coined the term desirable difficulties in the 1990s for conditions that slow practice yet improve long-term retention. A 2019 paper in Nature Communications by Robert Wilson and colleagues modeled learning in neural networks and found the fastest progress when the error rate sat near 15 percent. Tasks that were too easy or too hard slowed learning in the models.
More from Andrew Huberman
