No free lunch from deep learning in neuroscience: A case study through models of the entorhinal-hippocampal circuit

Rylan Schaeffer, Mikail Khona, Ila Fiete

Advances in Neural Information Processing Systems Accepted

December 2022

Neuroscience Deep Learning Grid Cells NeuroAI

PDF Code Poster Slides Video Tweeprint

Abstract

We challenge the notion that deep learning offers a free lunch for neuroscience. Using deep network models of the entorhinal-hippocampal circuit, we show that grid-like units only emerge under biologically invalid hyperparameter choices, and lack key properties of biological grid cells.

Summary

Deep learning models of the brain don't automatically provide scientific insight without careful analysis.

Summary

The promises of deep learning-based models of the brain are that they (1) shed light on the brain’s fundamental optimization problems/solutions, and/or (2) make novel predictions. We show, using deep network models of the MEC-HPC circuit, that one may get neither.

Key Findings:

1. Grid units rarely emerge: Of >11,000 networks trained, most learned to path integrate but <10% exhibited possible grid-like units. Path integration does not generically create grid units.

1. Specific (problematic) choices required: Grid units emerge only under specific supervised target encodings that require biologically incorrect assumptions about place cells.

2. Missing key properties: Even when grid-like units emerge, they lack key properties of biological grid cells (multiple modules, specific ratios between modules).

3. Extreme sensitivity: Grid-like unit emergence is highly sensitive to hyperparameters—changing place cell width from 12cm to 11cm or 13cm eliminates grid emergence.

Takeaway for NeuroAI: It is highly improbable that a path integration objective for ANNs would have produced grid cells as a novel prediction, had grid cells not been known to exist. Deep learning does not offer a free lunch for neuroscience.