Beyond the Manifold Hypothesis: Hybrid Spectral Parameterizations for Flow Matching
This analysis identifies two main drivers for differences in source-- data signal-to-noise ratio, and the information bottleneck induced by the neural architecture, and introduces new spectral hybrid parameterizations that adapt across time and data covariance directions.
Ségolène Martin, Anne Gagneux, Quentin Bertrand et al.
· 0 citations