Project information
mosAIc: The hidden geometry of spiking neural networks (dr. Dominik Dold)

Modern artificial intelligence (AI), powered by artificial neural networks (ANNs), is facing rapidly increasing energy demands. As a brain-inspired alternative, spiking neural networks (SNNs) promise far greater energy efficiency by mimicking how the brain processes information using sparse, time-based electrical pulses (spikes). However, the computational capabilities of SNNs, and how to exploit their temporal dynamics, remain poorly understood. This knowledge gap impedes their widespread adoption and leaves a key question unanswered: Can SNNs match or outperform ANNs while using significantly less energy?

mosAIc introduces a new theoretical concept to address this challenge: causal pieces (unrelated to causal inference / causality). A causal piece is a region in the input and parameter space of an SNN where the output spikes are caused by the same subnetwork - decomposing input and parameter space into a mosaic of pieces. Inspired by affine pieces used to analyse ANNs, mosAIc’s central hypothesis is that the number and geometry of causal pieces quantify the expressivity of SNNs. This enables a unified analysis of SNNs across neuron models, and a principled comparison between SNNs and ANNs.

The project pursues four key objectives, moving from theory to practical application:
(i) Develop a general, model-agnostic definition of causal pieces applicable across spiking neuron models.
(ii) Identify computational properties that can be predicted from causal piece structure, such as expressivity, generalisation, and energy demands.
(iii) Use causal pieces to guide SNN design, e.g., to improve parameter initialisation strategies and training methods.
(iv) Benchmark SNNs against ANNs on real-world edge-AI tasks - specifically onboard AI tasks - providing the first principled comparison based on causal and affine pieces.

To maximise impact, the project will release open source software and host a public online competition on optimising the causal piece structure of SNNs.