Title: Compositional cognition: learning a model of the world from its parts
Abstract: Humans can perform remarkably complex tasks, such as flying an aircraft or playing a violin concerto. The collection of mechanisms that underlie task-level performance ("executive functions") and their neural implementation in the prefrontal cortex have been extensively investigated by psychologists and neuroscientists. However, we know remarkably little about how new tasks are learned. This is a pressing problem, because despite recent advances in machine learning, researchers are currently unable to build intelligent systems that learn to perform multiple complex tasks in series (e.g. successive Atari games) without resetting network parameters. My talk will focus on the challenges of understanding task learning in humans, and describe recent work that has suggested that complex tasks can be best solved when broken down into their constituent parts (compositional learning). I will illustrate with examples from tasks involving navigation, visual categorisation, and value-guided learning in novel environments.