Samuel A. Ocko

This paper shows that hexagonal grid cell patterns emerge in neural networks trained for path integration and develops methods to analyze these networks for mechanistic insights.

The paper shows that medial entorhinal cortex uses Kalman filtering to integrate path integration and landmarks, while visual cortex codes position retrospectively and retrosplenial cortex is intermediate.

Imposing an optic-nerve-like bottleneck in a deep convolutional neural network trained on image recognition causes antagonistic center-surround retinal receptive fields and orientation-tuned primary visual cortex representations to emerge simultaneously.

The paper develops a neural attractor model that explains how animals construct spatial maps using self-motion and landmark cues, showing emergent elasticity in grid cell firing patterns.

Optimizing linear and rectifying nonlinear retinal encoding models on natural movies under a fixed firing-rate budget yields ON and OFF midget- and parasol-like ganglion cell types matching primate retina organization.