Generative Models

I recently had the pleasure to give a 60-minute fireside chat on Distributional Autoencoders Know the Score, hosted by Ploutos. Recording (Ploutos World) Clubhouse room link Paper page (with links)

I recently posted a new preprint titled Energy-Tweedie: Score meets Score, Energy meets Energy on ArXiv.

We first extend the classic Tweedie’s formula to a wider range of distributions, then derive a new identity that links the (Stein) score under a generalized Gaussian (energy model) noise to the gradient of a matched Energy Score (a scoring rule), connecting two seemingly disparate approaches to generative modeling and serving as an analog to the classic Tweedie’s formula. Among other things, this represents a novel score matching estimator, as well as enables energy-score-based denoising models to be plugged into standard diffusion pipelines, but with heavy‑tailed noise.

For the Distributional Principal Autoencoder (DPA), we prove an exact identity linking the data score to the geometry of the encoder level sets, and show that any latent coordinates beyond the data manifold become completely uninformative. This means that the DPA learns nonlinear manifolds shaped locally by the data density, with a clear, testable dimensionality criterion — conditional independence, indicating it can be considered a natural nonlinear analog to PCA.

I recently helped present our work (together with Prof. Gašper Beguš, paper) on the usage of deep generative models for discovering the building blocks of sperm whale communication at the Decoding Communication in Nonhuman Species symposium hosted at the Simons Institute for the Theory of Computing, UC Berkeley:

We propose a novel methodology - Causal Disentanglement with Exteme Values (CDEV) - to identify representations learned by GANs. When trained on raw whale communication, it finds - for the first time - specific acoustic attributes that might serve as carriers of meaning.