How to Model the Mind Simultaneously Across the Computational, Algorithmic, and Neural Levels

​​In the history of neuroscience and psychology, we can, for the first time, discover complex algorithms of intelligence in concrete, computational terms. I will present my lab’s work, which advances these efforts by focusing on visual cognition, uncovering the computational logic and intermediate representations that transform images into rich representations of objects and scenes that we can think about and plan with. The overall theoretical core of this work is that visual cognition is fundamentally about building and manipulating ‘structure-preserving representations’ (SPRs) of the physical world, going beyond the task-optimized statistical representations featured in standard deep neural networks. I’ll present multilevel formulations of this core theory of SPRs, making contact with empirical measurements across levels of analysis, from dense psychophysics to single-cell electrophysiology. I’ll highlight our work in the domains of object perception, intuitive physics, and goal-driven attention. These studies blur the boundaries between currently divergent modeling approaches of cognitive science (probabilistic/connectionist/dynamical systems), provide accounts of neural mechanisms that are simultaneously more interpretable and predictive than alternatives, and offer a way to synthesize task optimization and structure preservation.