Steering AI Representation Geometry Improves Brain Alignment

Recent studies have shown that the alignment between artificial and biological neural networks is often asymmetric: AI model representations are better at predicting brain activity than brain activity is at predicting AI representations. This asymmetry raises questions about the underlying factors contributing to this imbalance. This paper investigates whether the geometric structure of representations plays a role in achieving more balanced, bidirectional alignment. The researchers hypothesized that by systematically steering the representational geometry during model training, they could influence this bidirectional alignment. They developed a computational framework that combines spectral regularization with bidirectional predictivity analyses. As a proof of concept, they applied this framework to self-supervised contrastive vision models. The results showed that steering the spectral geometry of the learned representations substantially increased "reverse predictivity" (how well neural responses predict model representations) while only modestly reducing "forward predictivity" (model predicting neural responses). This led to a 55% relative improvement in overall bidirectional predictivity. These improvements were linked to reduced effective dimensionality and a reorganization of the shared representational subspace, where forward and reverse predictivity became more symmetric. This work highlights that representational geometry is a key factor in modulating the alignment between artificial and biological intelligence.