Title:  Establishing Trust in Neural Networks with Representation Shifts

Committee: 

Dr. AlRegib, Advisor    

Dr. Dyer, Chair

Dr. Kira

Abstract: The objective of the proposed research is to improve the trust in neural networks with second-order representation shifts. While neural networks are effective in modeling complex data dependencies, human trust remains a major obstacle for industrial deployment and network predictions are frequently met with sincere skepticism. In this work, I propose establishing trust by integrating robustness, consistency, and uncertainty-awareness natively within the neural network paradigm. For this purpose, I extract statistics from shifts of neural network representations (second-order representations) and predict the learning difficulty during deployment. The approach is modular and generalizable to applications where deterministic neural networks are deployed.