The 85% Rule for Learning

Learning, it seems, is optimized for both humans and machines when we succeed around 85% of the time. From a recent paper by Wilson, Shenhav, Straccia and Cohen:

In many situations we find that there is a sweet spot in which training is neither too easy nor too hard, and where learning progresses most quickly. […] For all of these stochastic gradient-descent based learning algorithms, we find that the optimal error rate for training is around 15.87% or, conversely, that the optimal training accuracy is about 85%.

If you’re always successful, it’s hard to know what to improve. If you constantly fail, you won’t learn what works. Only when we have a mixture of success and failure can we draw a contrast between good and bad strategies.

These findings agree closely with the 80% success rate found by Barak Rosenshine in his study of successful classrooms, despite coming from a completely different theoretical background:

In a study of fourth-grade mathematics, it was found that 82 percent of students’ answers were correct in the classrooms of the most successful teachers, but the least successful teachers had a success rate of only 73 percent. A high success rate during guided practice also leads to a higher success rate when students are working on problems on their own.