Title:  Contrastive Reasoning in Neural Networks

Committee:

Dr. Ghassan AlRegib, ECE, Chair, Advisor

Dr. Mark Davenport, ECE

Dr. Mark Riedl, CoC

Dr. Larry Heck, ECE

Dr. Eva Dyer, BME

Abstract: The objective of the dissertation is to rethink the inductive nature of reasoning in neural networks by providing contextual explanations to a network's decision and addressing the network's robustness capabilities. Neural networks represent data as projections on trained weights in a high dimensional manifold. The trained weights act as a knowledge base consisting of causal class dependencies. Inference built on features that identify dependencies within this manifold is termed as inductive feed-forward inference. This is a classical cause-to-effect inference model that is widely used because of its simple mathematical functionality and ease of operation. Nevertheless, feed-forward models do not generalize well to untrained situations. To alleviate this generalization challenge, we use an effect-to-cause inference model that falls under the abductive reasoning framework. Here, the features represent the change from existing weight dependencies given a certain effect. In this dissertation, we term this change as contrast and the ensuing inference mechanism as contrastive inference. The dissertation tackles contrastive reasoning and inference in three phases: 1) We formalize the structure of contrastive reasons as answers to questions of the form ·why P, rather than Q?' and provide a simple mechanism to extract contrastive reasons from any pre-trained neural network, 2) We analyze our contrastive reasons by creating a taxonomy of evaluation, and 3) Use contrastive reasons to make robust and generalizable decisions. We provide representational and explanatory insights using our contrastive reasoning scheme for diverse visual applications including: 1) Visual cognition-based human attention, 2) P robabilistic modeling of causal and context features, 3) Active learning paradigm of machine learning, and 4) Meta-representations within neural networks.