CSL Master Calendar

Title: Neural Diffusion Intensity Models for Point Process Data

Abstract: This talk will present a flexible, generative approach to modeling and forecasting over-dispersed point process data. We model these as doubly stochastic Poisson processes (Cox processes) whose intensity follows a stochastic differential equation (SDE). We represent the SDE drift function with an expressive neural network, and develop a novel variational approach to fit this to data. Using tools from the theory of enlargement of filtrations, we show that a variational family indexed by the drift neural network is rich enough that parameter fitting by evidence lower bound (ELBO) maximization coincides with likelihood maximization. In addition to learning the prior drift this way, our approach also learns a posterior correction drift, modeled with an encoder architecture that takes as input a varying-length point process observation. Our novel architecture amortizes a smoothing posterior over paths, and avoids the heavy MCMC simulations that traditional expectation-maximization (EM) approaches require for posterior inference.

Bio: Harsha Honnappa is an Associate Professor in the Edwardson School of Industrial Engineering at Purdue University, where he directs the Stochastic Systems Lab. He is an applied probabilist with interests in the modeling and analysis of stochastic systems, theoretical statistics, stochastic optimization and control. His research is supported by grants from the National Science Foundation, including an NSF CAREER award, the Office of Naval Research, the Purdue Research Foundation, and through the Edwardson School of Industrial Engineering Frontiers awards. He is currently Area Editor for Stochastic Models for Operations Research Letters, and Associate Editor at Operations Research and Queueing Systems Journals.

Location & Time: CSLB02, February 18, 3-4PM
Reception in CSL154 at 2:30PM