Learning is optimized when we fail 15% of the time

The study argues that there is a measurable sweet spot for learning. In simple tasks with clear correct and incorrect answers, the best learning rate appears when failure happens about 15% of the time, or when responses are correct 85% of the time.

The authors tested this idea with machine-learning experiments and found the same pattern in previous animal-learning studies. They suggest it may also apply to perceptual learning in humans, but they limit the claim to simple tasks and do not extend it to general school grades.

Reading notes

#

• Learning is strongest when a task is challenging enough to leave room for improvement, but not so hard that progress stops.
• In the experiments, computers learned fastest when they reached 85% accuracy on two-choice tasks.
• The 85% rule also matched patterns found in earlier animal-learning studies.
• The authors think the idea may fit perceptual learning, such as learning to distinguish tumors from non-tumors in images.
• The study does not claim that students should aim for a B average, because it only tested simple tasks with clear right and wrong answers.
• The authors say the work may help education research on more complex forms of learning.