Complementary Learning Systems in Brains and Machines

Striking findings of the effects of brain damage on human memory have motivated a theory called Complementary Learning Systems theory (McClelland, McNaughton & O'Reilly, 1995) in which distinct brain areas work together to support both the gradual acquisition of structured knowledge and the ability to learn new things rapidly without interference with knowledge previously acquired. In a sense, today's AI systems also have complementary learning systems too. In this talk I will describe the complementary learning systems in the brain as well as how they relate to transformer based learning systems, using the modern Hopfield network (Krotov & Hopfield, 2016, 2021; Ramsauer et al 2020) as a bridging framework. I will suggest steps we are taking in my lab to build models that capture some of the properties of brain-like learning systems and suggest why we think these systems may help overcome some of the limitations of today's AI systems.