Reverse-engineering recurrent neural network solutions to a hierarchical inference task for mice

Rylan Schaeffer, Mikail Khona, Leenoy Meshulam, International Brain Laboratory, Ila Rani Fiete

Advances in Neural Information Processing Systems Accepted

December 2020

Neuroscience Recurrent Neural Networks Interpretability

PDF Poster Slides

Abstract

We reverse-engineer recurrent neural networks trained on a hierarchical inference task that mice perform, revealing interpretable dynamics and representations that provide hypotheses for neural implementation.

Summary

Reverse-engineering RNN solutions to understand hierarchical inference in mice.

Summary

This work investigates how recurrent neural networks solve a hierarchical inference task that mice perform in the International Brain Laboratory experiments. By training RNNs to perform the same task and reverse-engineering their solutions, we gain insights into potential neural implementations of hierarchical inference.

Key Contributions:

1. Training RNNs to match mouse behavior on hierarchical inference tasks
2. Reverse-engineering the learned solutions to identify interpretable dynamics
3. Generating testable hypotheses for neural recordings