Blood and Modularity: A Brain-Inspired Approach to Functional Disentangled Neural Representations

Authors: Nicolas Dundov, Dan Goodman

Presenting author: Nicolas Dundov

Presentation type: Poster at SNUFA 2026 online workshop (4-5 Nov 2026)

Abstract

Intelligent systems in real environments must process multi-modal stimuli in real time. At the neural network level, this requires coordinating complex interactions across an expansive state space. Given the combinatorial explosion in the number of potentially relevant interactions between inputs, it quickly becomes intractable to compute every possibility. Given the brain is a resource-constrained system, it must have evolved efficient mechanisms for dealing with this problem. A trait of particular interest is modularity: The emergence of separable representations within the network from distinct or entangled inputs. Purely functional modularity allows a system to break down complex functions into separable sub-functions across the state space without having to compute every representation, resulting in an energy efficient network with fewer and more focused activations. It may also facilitate transfer learning across solutions that generalise beyond the training distribution. We developed a novel network architecture featuring an energy allocator inspired by blood distribution mechanisms in the brain during periods of intense focus (fig 1a). In our model, the allocator network, which is jointly trained with the primary network, decides how much energy each spatial region should receive (with a fixed total budget). The less energy a region has, the noisier its outputs. This encourages the network to align spatial regions with functional modules. We evaluated this model on a paired KMNIST/MNIST modular classification task. To understand what influences the emergence of modularity in a multi-layer perceptron (MLP), we conducted a parameter grid search (fig 1b). Changing the total energy available for distribution had a significant effect, where excess energy led to an exponential decay in modularity, suggesting limited resource allocation enhances emergence. The network successfully achieves functional modularity and high classification accuracy across a fully connected MLP architecture, isolating distinct representations without structural partitioning. This could both provide a scalable framework to encourage the emergence of functional modularity in Artificial Neural Networks and potentially explain its emergence in the brain. This blueprint could be translated into varied network configurations and sizes benefiting specialised, energy efficient neuromorphic computing systems.

Are you the author and do you want to add some related links to this page? If so, check out our guide to adding related links and videos.