Automatic design of brain-like neural models through evolution

What is it about?

Our brains process information using populations of neurons that activate together across space. These dynamics can be modelled using "dynamic neural fields" — mathematical descriptions of how these populations of activity rise, spread, and fade over time. These models are useful for understanding things like memory, attention, and decision-making. The problem is that building these models has always required a human expert to hand-design them: choosing how many fields to use, how they connect, and tuning many parameters by trial and error. In this paper, we introduce NEAT-DNFs, a method that lets these models design themselves through a process inspired by biological evolution. Starting from the simplest possible model, our system automatically tries small variations, keeps the ones that work better at a given task, and gradually grows more complex structure — but only when the task actually requires it.

Featured Image

Photo by Logan Voss on Unsplash

Photo by Logan Voss on Unsplash

Why is it important?

Until now, building these brain-inspired neural models has depended entirely on manual design by experts. Our approach removes that bottleneck: it automatically discovers both the internal workings of each model component and how the components should be connected, without any manual tuning.