BraiNCA: Brain-Inspired Neural Cellular Automata for Morphogenesis and Motor Control
This presentation explores how brain-inspired architectural principles—attention-based message routing, long-range connectivity, and functional topology—transform neural cellular automata into more robust, adaptive, and sample-efficient distributed systems. Through rigorous experiments in morphogenetic pattern formation and decentralized motor control, the work demonstrates that these biological design principles yield quantifiable performance gains: up to 54% faster learning in morphogenesis and improved robustness in motor tasks through somatotopic organization.Script
Most artificial neural systems ignore a fundamental insight from neuroscience: real brains don't just connect neighbors—they wire distant regions together with attention-driven routing and functional specialization. What happens when we build that architecture into distributed cellular automata?
BraiNCA extends standard neural cellular automata with three biological principles. Each cell integrates information from both its immediate neighbors and a sparse set of distant nodes, all modulated by separate attention mechanisms. The architecture works on arbitrary graphs, not just regular grids, enabling functionally specialized regions like those found in motor cortex.
First, the researchers tested whether these principles accelerate self-organization in pattern formation tasks.
In decentralized pattern formation, cells on a 16 by 16 grid must self-organize into a target image from random initialization, with no global positioning information. Adding long-range connections to a standard 3 by 3 neighborhood reduced training episodes by 31%. Combining larger neighborhoods with long-range wiring achieved a 2.19 times speedup—357 fewer episodes to reach 98% accuracy.
The second benchmark tests whether topology matters for robust sensorimotor coordination.
In the Lunar Lander task, a grid of cells must collectively control thruster actions through decentralized voting. The T-shaped topology, which partitions action zones to mimic motor map organization in biological brains, outperformed both standard grids and naive long-range wiring. Success rates jumped from 84% to 88%, and learning accelerated significantly. Surprisingly, adding long-range connections without functional regionalization actually hurt performance—connectivity alone isn't enough; the wiring pattern must respect task structure.
BraiNCA proves that brain-inspired design principles—selective attention, strategic long-range wiring, and functional topology—aren't just biological curiosities. They're quantifiable levers for building faster, more robust distributed intelligence. Visit EmergentMind.com to explore the full paper and create your own research videos.
