New SRF method decodes interpretable dimensions from brains and AI

Researchers Florian Mahner and colleagues at the National Institutes of Health and MPI CBS have introduced SRF (Similarity-Based Representation Factorization), a computational method that extracts interpretable, low-dimensional embeddings from similarity matrices. These matrices are common across brain imaging, behavioral experiments, and AI model activations, but existing methods often produce opaque or incomplete dimensions. SRF addresses this by decomposing similarity data into non-negative components that are directly interpretable, even when the original data is sparsely sampled or incomplete. The method was validated on simulations and multiple real-world datasets spanning neural recordings, human judgments, and deep network representations.

SRF consistently recovered dimensions that matched those from task-specific models and predicted independent behavioral outcomes. It also outperformed standard similarity-matrix comparisons in exploratory analysis and confirmatory hypothesis testing, offering higher statistical power. This makes SRF a versatile tool for bridging disciplines: neuroscientists can understand what brain regions encode, psychologists can map cognitive features, and AI researchers can peek inside model internals. The paper, available on arXiv, positions SRF as a general-purpose method for revealing the core dimensions underlying representations across brains, behavior, and machines.