The Geometry of Generative Cognition

Intelligence requires a finite neural system to generate and flexibly perform a near boundless repertoire of tasks. Humans perform a wide array of different tasks, learning new tasks from a few examples or from instructions alone. Despite a growing understanding of how the brain performs specific well-trained behaviors, we lack a mechanistic model of how many different tasks are learned and performed at a timescale that precludes significant synaptic weight changes. In this talk, I will present a new model that uses the geometric structure of transformation groups to rapidly infer and perform tasks, even performing completley novel tasks. I will relate the model to recent experimental work in my lab studying the compositional reuse of subspaces and the geometry of task representations in the brain.