Increasing Bidirectional Alignment via Spectral Regularization

A new computational framework improves bidirectional representational alignment between biological and artificial neural networks by steering representational geometry. Using spectral regularization on self-supervised vision models, the method achieved a 55% relative improvement in reverse predictivity.