What hidden dimensions shape how brains and AI see the world?

Neuroscientists and AI researchers study representations by comparing how similar different stimuli are, but this tells you little about the underlying dimensions that actually drive perception and learning. A new technique called Similarity-Based Representation Factorization (SRF) extracts low-dimensional, interpretable embeddings from similarity data across neural recordings, behavioral tests, and AI models. The method works even with incomplete data and recovers dimensions that match task-specific models, predict behavior, and boost statistical power—useful for anyone trying to understand what representations actually encode.