Tag
This paper isolates the irreducible excess in denoising score matching training loss, showing it equals the trace of the Fisher-Rao metric along diffusion trajectories, connecting variational principles, latent space geometry, and information theory.
This paper introduces Fisher8, an output-layer gradient correction that uses Fisher geometry instead of Euclidean geometry to stabilize neural heteroscedastic regression, improving uncertainty calibration and likelihood-error tradeoffs.