NeuroAI: interpreting deep neural networks through the lens of biological processing

Deep neural networks can now solve many visual tasks, and often do so with internal representations that resemble those of the brain. Yet without principles for interpreting what these networks compute, we understand neither biological nor artificial computation. On the one hand, we use machine learning tools to capture the computations performed within populations of recorded neurons, and develop methods to link them to cellular and circuit mechanisms. On the other hand, we look to how artificial systems solve complex tasks to understand how computations might be structured in biological systems.

Relevant publications and presentations

[SFN 2025] Functional mapping of neurons in primary visual cortex using computational barcodes, presented by Isabel Fernandez (co-authors Ján Antolik, Dan Butts) [Poster]
Vafaii H, Yates JL, Butts DA (2023) Hierarchical VAEs provide a normative account of motion processing in the primate brain. bioRxiv 2023.09.27.559646 (presented at NeurIPS 2023). [NeurIPS, Twitter]
[SFN 2023] Hierarchical VAEs provide a normative account of motion processing in the primate brain, presented by Hadi Vafaii [Poster]
[SFN 2023] Biologically constrained deep neural networks to parse visual computations being performed in the primary visual cortex, presented by Matt Jacobsen [Poster]