Computer science: An optimum difficulty level for learning

Learning occurs most quickly when the difficulty of the training is adjusted to keep the learner’s accuracy at around 85%, reports a computational study in Nature Communications. The findings could aid the development of a theory to identify the optimal settings for maximizing the rate of learning.

Researchers have debated questions around the optimum conditions for teaching for a number of years. However, it is unclear why a particular difficulty level may be beneficial for learning and what that optimal level might be.

Robert Wilson and colleagues examined whether the level of difficulty of training affects the rate of learning. The authors derived the conditions for optimal learning for a set of algorithms in a series of binary classification tasks. Here, learning-based algorithms had to classify ambiguous stimuli into one of two classes (for example, the direction of travel of a small number of coherent dots in a random pattern). They found that the optimal error rate was around 15.87%, or conversely, that training progressed most quickly when the algorithm’s learning accuracy was around 85%. They also found that training at the optimal accuracy proceeded faster than training at a fixed difficulty.

When applied to artificial networks used in artificial intelligence and a model from computational neuroscience that is thought to describe human and animal perceptual learning, the authors found that the optimal rate of learning followed their ‘Eighty Five Percent Rule’.