Activity-Dependent Plasticity in Morphogenetically-Grown Recurrent Networks

A team of researchers has published a novel study on creating more adaptable AI for robotics. The paper, 'Activity-Dependent Plasticity in Morphogenetically-Grown Recurrent Networks,' introduces a co-evolutionary method where AI controllers are not just grown from a compact genome but also have their plasticity rules—how they learn and change—encoded within that same genome. This allows the networks to adapt their own connections after being deployed. The team tested a staggering 5 million+ configurations across standard control benchmarks like CartPole and Acrobot to characterize the effectiveness of different learning rules.

Their key finding is that anti-Hebbian plasticity, which weakens connections between simultaneously active neurons, significantly outperformed traditional Hebbian plasticity ('neurons that fire together, wire together') for competent networks. The effect size was substantial, with a Cohen's d between 0.53 and 0.64. Crucially, when they let evolution discover the best plasticity rules independently, it consistently converged on anti-Hebbian patterns, validating the finding. The research shows that without this built-in adaptability, networks suffer a 'regret'—losing 52-100% of potential performance improvement. Morphogenetically grown networks showed 2 to 6 times higher regret than random networks when deprived of plasticity, proving the unique adaptability of the developmental approach.