Reproduce “Functional differentiations in evolutionary reservoir computing networks” on OpenResearch

About this paper

Yutaka Yamaguti, Ichiro Tsuda

An Evolutionary Reservoir Computer (ERC) autonomously develops functional differentiation in its internal units when optimized for separating combined sensory inputs, achieving around 90% classification accuracy for both spatial and temporal patterns. This computational model, developed by researchers at Fukuoka Institute of Technology and Chubu University, demonstrates that specialized neurons emerge through structural evolution, including strengthened feedforward and weakened feedback connections.

Key insights

• Functional differentiation, characterized by individual neurons specializing in either spatial or temporal information, emerges robustly in the output layer of ERCs trained on separation tasks.
• Evolutionary optimization of the internal network structure leads to a strengthening of feedforward connections from input to output layers and a weakening of feedback connections, consistent with some biological brain network architectures.
• The specific task demand dictates the emergent neural coding strategy: separation tasks induce functional differentiation, while combination tasks foster mixed selectivity where neurons respond to multiple input features simultaneously.

Summary by alphaXiv.