Abstract: As data-driven machine-learnt algorithms become the technology of choice in an increasing array of applications, the research community recognizes the urgency of addressing shortcomings such as the lack of robustness (e.g., against adversarial examples and distribution shifts) and fairness (e.g., caused by bias in the training data).  In this talk, we present two architectural insights, each based on a shift of perspective from the state of the art.
1) Software Architecture: We view the standard end-to-end paradigm for training DNNs, which does not provide explicit control over the features extracted by intermediate layers, as a fundamental bottleneck in the design of robust, interpretable DNNs. Motivated by ideas from communication theory (learning matched filters) and neuroscience (neuronal competition), we propose adapting the training and inference framework for DNNs to provide more direct control over the shape of activations in intermediate layers. Preliminary results for the CiFAR-10 image database indicate significant gains in general-purpose robustness against noise and common corruptions, as well as against adversarial perturbations.  We hope these results motivate further theoretical and experimental investigations: variants of the ideas we propose apply, in principle, to any DNN architecture or training model (supervised, unsupervised, self-supervised, semi-supervised).
2) Social Architecture: We view unfairness in DNNs resulting from data bias as a symptom of the unfairness and bias in the society from which the data is extracted.  In an approach that is complementary to existing research on enhancing fairness during training and inference, we propose a framework for sequential decision-making aimed at dynamically influencing long-term societal fairness via positive feedback.  We illustrate our ideas via a problem of selecting applicants from a pool consisting of two groups, one of which is under-represented, and hope that our results stimulate the collaboration between policymakers, social scientists and machine learning researchers required for real-world impact.

Bio: Upamanyu Madhow is Distinguished Professor of Electrical and Computer Engineering at the University of California, Santa Barbara.  His current research interests focus on next generation communication, sensing and inference infrastructures, with emphasis on millimeter wave systems, and on fundamentals and applications of robust machine learning. Dr. Madhow is a recipient of the 1996 NSF CAREER award, co-recipient of the 2012 IEEE Marconi prize paper award in wireless communications, and recipient of a 2018 Distinguished Alumni award from the ECE Department at the University of Illinois, Urbana-Champaign. He is the author of two textbooks published by Cambridge University Press, Fundamentals of Digital Communication (2008) and Introduction to Communication Systems (2014).  Prof. Madhow is co-inventor on 32 US patents, and has been closely involved in technology transfer of his research through several start-up companies, including ShadowMaps, a software-only approach to GPS location improvement which was deployed worldwide by Uber.