Motifs, Modules, and Mutations: Building Brain-like Networks

Networks exist in various forms in living organisms, including neural, metabolic, gene regulatory, and signal transduction networks. While Artificial Neural Networks (ANNs) are inspired by the first, studying the ‘biochemical connectionism’ of other networks may provide insights into how network structure affects function. Traditional network science metrics like degree distribution and average path length can describe these networks, but they don’t fully understand their functionality. Systems Biology has made progress in analysing the microstructure of biological networks and their relationship to function using network motifs. Neural networks also exhibit larger-scale repeated structures, such as cortical micro-columns, which may form functional modules. Repeated structures, from small motifs to larger modules, can benefit networks.

Neuroscientists highlight the gap between biological neural networks and ANNs in this regard. Biological networks are rich in structure, with a developmental process unfolding genetic information into the final network. In conventional feedforward ANN architectures, the structure is predetermined and fixed, and weights are trained. Training procedures like backpropagation usually don’t result in repeating network structures. When network structure is the result of an evolutionary algorithm, modularity doesn’t typically arise. Modular networks tend to disappear over evolution due to mutations that improve task fitness and disrupt modularity. However, preserving modularity provides benefits such as robustness, generalisation, interpretability, and evolvability, even if it reduces fitness for a specific task.

In this context, I will discuss our research on different approaches to neuroevolution. We use genetic algorithms and novelty search to create networks of artificial (spiking) neurons. Motifs inspired from the peripheral nervous system serve as components for larger spiking neural microcircuits targeted at fault-tolerance applications. A developmental graph cellular automata (DGCA) based on models of evolutionary and developmental processes is designed to produce far-from-random motif distributions similar to what is observed in biological organisms.

Workshop main page