Mo Shakiba · NeuroAI researcher

Mo Shakiba

I’m Mo, a NeuroAI researcher.


On: 2026-07-27 Post

Neuronal Constraints Drive Superior Learning in RNNs

Most RNNs still start from a blank slate: random weights, no space, no wiring. Cortex doesn’t work that way! Neurons sit in physical space, wire economically, and organize into structured functional relationships. Using the MICrONS dataset, we built RNNs whose weights, spatial embedding, and communicability constraints come directly from that measured cortex.

Across eleven model variants and three cognitive tasks, function-derived weight initialization gave the largest gain in task performance, real neuronal coordinates added a robust secondary boost, and communicability-aware regularization refined the learned solution more selectively. The resulting networks also converge on brain-like organization—low entropy, modularity, small-worldness—and stay trainable under a positive-only-weight constraint that collapses randomly initialized models to chance.

The article1 is now up on arXiv—a continuation of the project we first introduced through the Impact Scholars program2. There’s also a companion blog post3, and a few have already written about it45. Below is the video overview that covers the paper comprehensively.