Intra-modal dispersion modulates cross-modal neural convergence

A recent study presents a novel approach employing the Generalized Procrustes Algorithm to assess intra-modal representational convergence at the level of single stimuli within neural networks. The researchers focused on vision models with varying training goals, choosing stimuli based on their alignment degree. Their findings indicate that intra-modal dispersion significantly influences the alignment between vision and language models, providing insights into how specific stimuli generate convergent representations across different networks. This research expands on the idea that neural networks achieve similar representations by learning shared environmental structures, while also clarifying the impact of individual stimuli. The study is available on arXiv as paper 2604.21836.