Modulating Cross-Modal Convergence with Single-Stimulus, Intra-Modal Dispersion

TL;DR

This paper introduces a method to measure how individual stimuli influence the convergence of neural network representations across modalities, revealing that stimuli with consistent intra-modal responses enhance cross-modal alignment.

Contribution

It presents a novel single-stimulus measurement technique using the Generalized Procrustes Algorithm to analyze intra-modal and cross-modal convergence in neural networks.

Findings

Low intra-modal dispersion stimuli increase cross-modal alignment by up to two times.

The effect of stimulus dispersion on convergence is robust across different model pairings.

Measuring convergence at the single-stimulus level helps understand sources of neural alignment and divergence.

Abstract

Neural networks exhibit a remarkable degree of representational convergence across diverse architectures, training objectives, and even data modalities. This convergence is predictive of alignment with brain representation. A recent hypothesis suggests this arises from learning the underlying structure in the environment in similar ways. However, it is unclear how individual stimuli elicit convergent representations across networks. An image can be perceived in multiple ways and expressed differently using words. Here, we introduce a methodology based on the Generalized Procrustes Algorithm to measure intra-modal representational convergence at the single-stimulus level. We applied this to vision models with distinct training objectives, selecting stimuli based on their degree of alignment (intra-modal dispersion). Crucially, we found that this intra-modal dispersion strongly modulates alignment between vision and language models (cross-modal convergence). Specifically, stimuli with low intra-modal dispersion (high agreement among vision models) elicited significantly higher cross-modal alignment than those with high dispersion, by up to a factor of two (e.g., in pairings of DINOv2 with language models). This effect was robust to stimulus selection criteria and generalized across different pairings of vision and language models. Measuring convergence at the single-stimulus level provides a path toward understanding the sources of convergence and divergence across modalities, and between neural networks and human neural representations.