Arina Belova: Diffusion as an Alternative to Linear Methods in ANN-Brain Encoding Models

BCCN Berlin / Technische Universität Berlin

 

Abstract

Brain encoding, predicting neural activity from sensory stimuli, remains a open challenge in cognitive computational neuroscience, with applications ranging from fundamental research to brain-computer interfaces. Despite significant advances in stimulus representations using pre-trained artificial neural networks (ANNs), linear regression has remained the dominant encoding method for over a decade, constrained by its univariate assumptions, inability to capture full neural response distributions, and limited generalisability across subjects.

This thesis investigates whether Diffusion Models, a class of deep generative models designed to learn complex high-dimensional distributions, offer meaningful advantages over ridge regression for the brain encoding task. Using functional magnetic resonance imaging data from Natural Scenes Dataset (NDS) experiment, we address two central questions: (1) how does a diffusion-based encoder compare to the ridge regression baseline in terms of encoding performance, and (2) does joint multi-subject training improve encoding relative to subject-specific models?

To answer these questions, we conduct a series of ablation experiments examining data representation, model configuration, and generation regime to identify an best-performing diffusion encoding framework. We then benchmark this framework against ridge regression in the single-subject setting, before extending the comparison to multi-subject models trained jointly across participants.

Our results show that single-subject diffusion models underperform ridge regression in encoding accuracy, while multi-subject diffusion models achieve performance on par with regression — with both approaches remaining within the bounds of the fMRI noise ceiling.

We attribute this recovery in performance to the inter-subject information implicitly contained in the pooled multi-subject dataset, which the generative model is able to exploit in ways unavailable to single-subject specific baselines.

Our findings contribute to a growing body of work exploring large-scale, generalizable encoding models of the human brain, and speak to the broader potential of generative modelling in computational neuroscience.

 

Guests are welcome!

 

Additional information:

Master thesis defense

 

Organized by:

Dr. Wojciech Samek & Prof. Adrien Doerig