Cortical Blueprint from 12,000 Neurons Outperforms Standard RNNs on Decision Tasks

12,000 co-registered excitatory neurons from mouse visual cortex, complete with spatial coordinates, anatomical connections, and functional firing patterns, beat out standard RNN baselines on three cognitive decision-making tasks.

The work, from the Machine Intelligence from Cortical Networks (MICrONS) program, leverages dense calcium imaging co-registered with high-resolution electron microscopy from the same animal. That means each neuron's location in physical space, its wiring diagram, and its activity during visual processing are all known and tied together.

What the MICrONS Data Brings to RNNs

Instead of random or learned recurrent weights, the team initialized recurrent weights using the spatial coordinates, anatomical connectivity, and function-derived relationships from that population of 12,000 neurons. They also imposed communication-aware spatial constraints during training. The result: networks constrained by cortical structure and function consistently outperformed both fully unconstrained and partially constrained models.

Functional weight initialization alone provided the largest performance gain. Real spatial embedding (geometric constraints) delivered robust additional improvements across all three tasks. These aren't marginal wins on a single benchmark; they held across multiple cognitive decision-making tasks.

Functional Weight Initialization is the Biggest Winner

The paper reports that functional relationships between neurons - derived from their co-firing patterns - gave the strongest boost when baked into initial weights. That makes sense: we know cortical circuits are tuned through experience, and injecting that functional structure directly into the RNN's starting point seems to bypass a lot of random search.

Spatial embedding, i.e., constraining connectivity based on real 3D positions, added a smaller but consistent edge. Together, these two inductive biases - function and geometry - reduced the learning burden significantly.

Biologically Grounded Networks Develop Brain-Like Structure

These biologically grounded RNNs didn't just perform better; they also converged toward key organizational principles of biological computation. The networks developed low-entropy, modular, and small-world connectivity patterns. Even when recurrence was restricted to positive weights only (mimicking excitatory-only circuits), performance remained strong.

That's striking: constraining the network to a biological wiring rule didn't cripple it; instead, it pushed the network toward efficient, brain-like representations.

Next step: scaling this approach to larger cortical areas and testing on more complex sensorimotor tasks - and maybe porting the same inductive bias idea to transformer architectures.

Source: Harnessing cortical geometry, wiring, and function as inductive biases for recurrent neural networks
Domain: arxiv.org