The Cambridge ELLIS Unit Seminar Series holds talks by leading researchers in the area of machine learning and AI. Our next speaker for 2023 will be Dr. David Barber. Details of his talk can be found below.
Title: “Training Energy Based Models “
Abstract: Energy Based models are undirected models of the formp(x) \propto exp(-f(x)). These are very general models and provide an alternative to the currently popular directed models used in Diffusion Models and LLMs. However, training and sampling from these models is typically problematic. We consider an approach based on Denoising Score Matching and consider a Gibbs sampling to recover samples from the learned model. Remarkably, we show that there is an exact expression for the mean and covariance of the sampler. We will also highlight close connections to Diffusion Models.