The article explores the identifiability of quantized factors in disentanglement theory. It introduces a novel form of identifiability, termed quantized factor identifiability, under diffeomorphisms. The presence of independent discontinuities in the joint probability density of latent factors allows for the recovery of quantized factors. Theoretical foundations are laid out to develop algorithms for learning disentangled representations robustly. Empirical evidence from neuroscience and machine learning supports the concept of grid-like representations and their identification through independent discontinuities. The article proposes a method to align gradients with axes for identifying sharp density changes indicative of independent discontinuities. Experimental results demonstrate successful reconstruction of latent variables using this approach compared to existing methods like Linear ICA and Hausdorff Factorized Support.
לשפה אחרת
מתוכן המקור
arxiv.org
שאלות מעמיקות