Cortex Geometry and Wiring Improve RNNs on Cognitive Tasks

The standard recipe for training a recurrent neural network starts with random weight initialization — a blank slate that learns all its structure from data. A paper posted to arXiv in June 2026 by Mo Shakiba, Rana Rokni, Mohammad Mohammadi, and Nima Dehghani argues that neuroscience already has a better starting point, and that it has been sitting in the connectomics databases.

The team drew on the MICrONS connectomics resource, which combines dense calcium imaging co-registered with high-resolution electron microscopy from mouse visual cortex. From that dataset they pulled spatial coordinates and anatomical connectivity patterns covering nearly 12,000 neurons, then used that structure as an inductive bias: shaping initial weights and imposing what the authors call communication-aware spatial constraints during learning. The question was whether cortical organization — geometry, wiring, and functional relationships — could do useful work before gradient descent ever runs.

Tested across three cognitive decision-making tasks, the biologically constrained networks consistently outperformed both standard baseline models and partially constrained versions. The single largest performance driver turned out to be functional weight initialization, meaning that encoding how neurons relate to one another functionally mattered more than spatial constraints alone. The constrained networks also developed low-entropy, modular, and small-world organization — properties characteristic of biological neural circuits — and the approach held up even when recurrent weights were restricted to positive-only values.

The honest caveat is that mouse visual cortex is a specific source, and it is not yet clear whether these gains transfer to tasks or architectures far removed from that biological context. Three decision-making tasks is a real test but a narrow one, and the approach depends on access to the kind of rare, dense connectomics data that MICrONS represents — a resource most teams cannot easily replicate.

The forward-looking angle is about what happens as connectomics datasets become more available. If the organizational principles that make cortex computationally efficient can be packaged as reusable inductive biases, neuroscience data stops being purely a scientific resource and starts being an engineering one. The robustness under positive-only weights is a small but practical green flag for anyone thinking about applying this outside a lab setting.