Genomic bottleneck method evolves sparse neural networks for tough temporal tasks

A team led by Mani Hamidi has introduced a novel method for constructing neural networks that mimics how biological brains develop from a tiny genomic blueprint. In their paper, they use hypernetworks—neural networks that generate the weights of other networks—to learn a compressed generative process. This process produces the connectivity of a modular reservoir computer, a type of recurrent network specialized for processing temporal data. By combining curriculum-based meta-learning with modular reservoir computing, the approach yields sparse recurrent networks that are both efficient and robust.

The key innovation is the 'genomic bottleneck': the hypernetwork learns a compact representation (like a genome) from which the full connectivity of the reservoir is decoded. This biologically inspired pipeline allows the network to solve difficult temporal tasks, such as time-series prediction or sequence generation, with minimal additional training. The resulting networks are sparse—meaning fewer connections are used—which reduces computational cost without sacrificing accuracy. The work sits at the intersection of neural and evolutionary computing and suggests a path toward more efficient, brain-like AI systems.