Monday, October 10
12.10 - 1.00
CSE 4140

Speaker: Todd Coleman, Dept of Bioengineering, UCSD
Title: Feedback Information Theoretic Approaches to Active Learning and Causal Inference

Abstract:
In this talk, we discuss two topics of active learning and causal inference. In the former, we pose the problem of active learning from a team decision theory viewpoint - using the lens of "two agents cooperating to achieve a common goal". With this, we make a connection with reliable communication of an analog message point over a noisy channel with feedback as well as the theory of optimal transportation. We demonstrate an application of this framework within the context of a brain-machine interface. In terms of causal inference, we consider the problem of simultaneously observing many time series and attempting to identify the causal interactions amongst them. We re-visit Granger's notion of causality for two processes and generalize it to arbitrary modalities using a sequential prediction viewpoint. We show how under the expected log loss, the unique embodiment of Granger's viewpoint is the directed information. We then use ideas from coupled differential equations and extend this framework to networks using minimal generative model graphs. We demonstrate applications of this framework in terms of interacting neural processes whose causal interactions co-vary with brain function.

Coleman lab: http://coleman.ucsd.edu/