It turns out that the difficulty level of the problem that can be learned most efficiently is '85% correct answer level'


Liam Anderson

When learning something new, people and computers solve various exercises repeatedly and sometimes deepen their learning through mistakes. A research team in the United States has announced the research result that 'the experiment using a machine learning algorithm proved that the optimal difficulty level for learning is about 85%'.

The Eighty Five Percent Rule for optimal learning | Nature Communications

Learning is Optimized When We Fail 15% of the Time | UANews

Failing 15% of The Time Is The Best Way to Learn, if AI Is Anything to Go By

Problems that are solved in learning things will not lead to efficient learning if they are too difficult, but conversely they are too simple to be 100% correct. Although it was unclear how much difficulty was optimal for efficient learning, pedologists thought `` There should be something like a 'sweet spot' that optimizes learning efficiency somewhere '' I came.

Research team of

Robert Wilson and others, Associate Professor of Psychology at the University of Arizona A machine learning experiment was conducted.

When analyzing the growth rate of the computer with various conditions changed, when the computer solved the “problem that is wrong at 15% frequency, that is, the problem that can be answered correctly with 85% accuracy”, it is learned most efficiently understood. This figure is consistent with previous studies conducted using animals.



The research team says, “Theoretically, training with the optimal difficulty level of 85% has been shown to improve learning efficiency.” Wilson believes that the difficulty of the optimal problem found this time is likely to be applied in human perceptual learning where experience is important.

For example, in a training in which a radiologist distinguishes tumors from images, learning efficiency may be maximized by repeatedly solving problems with a correct answer rate of about 85% rather than problems that are too simple or difficult. there is. Of course, as the target person or computer continues to learn, it seems that the correct answer rate will improve even for the same question, so the difficulty level of the question needs to be adjusted frequently.

In fact, further research is needed to confirm that the results of this research can be applied to humans. However, as the starting point of the research, the difficulty level of “correct answer rate 85%” is well-balanced, and can be a useful experimental result for pedologists. Wilson said, “If you are studying in a class that is too easy, you will not be able to keep up with those who are struggling with moderate difficulty. I hope that discussions will begin on learning as well. '

by klimkin

in Science, Posted by log1h_ik