When tested with a continual visual learning task—say differentiating cats from houses or telling apart stop signs and school buses—deep learning algorithms equipped with selective resetting easily maintained high accuracy over 5,000 different tasks. Standard algorithms, in contrast, rapidly deteriorated, their success eventually dropping to about a coin-toss.
Called continual back propagation, the strategy is “among the first of a large and fast-growing set of methods” to deal with the continuous learning problem, wrote Drs. Clare Lyle and Razvan Pascanu at Google DeepMind, who were not involved in the study.