Neural circuits that grow while working: trained at just three developmental snapshots, a controller keeps visuomotor behavior at untrained ages
Synopsis
A recurrent neural controller must keep behaving while neurons are born, move, and change their intrinsic time constants during development; controllers optimized only on the initial or the mature substrate generally failed at other ages. The study feeds developmental time into a compact connectivity generator and optimizes the same generator at three, five, or fifteen developmental snapshots, testing it on two geometrically distinct substrates and two visuomotor tasks. Optimization at only three snapshots was often enough to keep performance high and consistent at trained and untrained ages, while the generated connectomes changed substantially with age and did not converge on a single sparsity or weight trajectory.
Interpretation
A single compact connectivity-generating program can carry a recurrent controller through continuous changes in its neural substrate and keep visuomotor behavior effective across development. Controllers optimized on the initial substrate declined sharply as development progressed, and controllers optimized on the mature substrate declined toward earlier ages, so each worked only near the age it was optimized on. On two geometrically distinct substrates (Fish and Ring) and two tasks (optomotor response, OMR, and phototaxis), developmentally aware controllers kept higher and more consistent fitness at trained and untrained ages across 11 evenly spaced developmental times, with 20 randomized testing trials at each.
Connectivity is not encoded age by age but computed by one generator from neuronal geometry and developmental time. Specifying every synapse at every age would be inefficient and a poor abstraction of biological development, where the genome supplies compact rules rather than a full wiring diagram. The generator is a compositional pattern-producing network (CPPN) from HyperNEAT that takes source and target neuron coordinates plus developmental time as inputs and outputs a weight and an expression gate for each allowed connection, while the substrate prescribes which neurons are born, where they sit, and how intrinsic time constants change.
Behavioral stability does not require an invariant connectome or one shared connectivity-level strategy. Sparsity, the fraction of excitatory connections, and mean nonzero weight showed no common trajectory across the four substrate-task combinations and three developmentally aware conditions; some controllers became sparser while others stayed dense. Developmental time had a reachable path to the weight or expression output in 58 of 60 developmentally aware runs and in all 12 best seeds, and a no-neurogenesis comparison showed some controllers relied on later-born neurons while the Dev. Aware-15pts best seeds on both OMR substrates stayed high-performing without them.
Perspective
The results rest on two abstract computational substrates, a Fish geometry inspired by zebrafish visuomotor organization and a Ring geometry, and on two visuomotor tasks; the substrate developmental trajectories are prescribed rather than evolved, and body mechanics and body growth are held fixed.
Developmental fitness was qualitatively less stable for phototaxis than for OMR, and several phototaxis controllers produced oscillatory trajectories that the authors hypothesize reflect an active scanning strategy; the results establish a computational principle rather than a detailed model of any particular animal's development.