Is Deep Learning Statistical Learning?

McClelland. , J.

Over 25 years ago, behavioral investigations of statistical learning in humans (from infancy to adulthood) and computational modeling work using simple recurrent networks introduced a revolution in thinking about the role that learning and sensitivity to statistical structure might play in shaping structured knowledge even in domains such as language, where a role for statistics had often been denied. More recently, deep learning architectures have achieved remarkable success in a range of complex cognitive capacities, including mastery of complex strategy games like Go and Chess, and demanding language processing tasks such as machine translation. Yet there remain limitations on what current models have achieved. Are the successes of deep learning examples of statistical learning? Are there limits on the kinds of cognitive capabilities that such methods can achieve? Will further exploration of these approaches lead to new insights into human learning abilities? My talk will review some of the successes as well as the limitations of current deep learning models in relation to human learning abilities and point to ongoing research directions I and others are pursuing to address the ways in which human abilities still exceed the achievements of deep learning.