SILO: Shaping and Sampling from Learned Embeddings

Abstract

Neural networks and other architectures progressively reshape the geometry of data space from data’s raw form to one that better suits high-level tasks.  This geometric perspective suggests that we can manipulate the shape of these embeddings to improve interpretability or performance on certain tasks, and that we can devise geometric algorithms to draw samples that resemble the data distribution.  Building on this intuition, in this talk, I will summarize recent efforts from my team to improve, accelerate, and fine-tune performance of diffusion and language models using the geometry of latent embeddings and the set of models themselves.

Bio

Justin Solomon is an associate professor of Electrical Engineering and Computer Science in MIT’s Computer Science and Artificial Intelligence Laboratory.  He leads the MIT Geometric Data Processing group, which studies problems at the intersection of geometry, large-scale optimization, and applications.