Task 001/002 - Neuro-inspired Algorithms and Theory

ABSTRACT:

Deep Neural Networks (DNNs) are driving a new industrial revolution. Yet, we currently have no way to make rigorous claims about DNNs' behavior as we are largely ignorant of their learned mechanisms, ie. the "black-box problem". Understanding these mechanisms is necessary to develop AI applications that are explainable, reliable and verifiable. A promising strand of research that I have been pursuing to understand DNNs has been borrowing insights and tools from neuroscience ---articulating hypothesis and testing them as if DNNs were another brain. My talk will review our recent progress made using such neuro-inspired approach to understand the failures and the successes of DNNs. Namely, I will speak about our efforts in understanding when and how DNNs recognize objects in novel viewpoints and image degradations.

 

BIO:

Xavier Boix is a postdoctoral researcher at the Department of Brain and Cognitive Sciences at MIT (since 2014) and is affiliated at Harvard (since 2017). He is also member of the multidisciplinary NSF research center for Brains, Minds and Machines (CBMM). Xavier obtained a doctorate from ETH Zurich (2014) and completed a postdoc at the National University of Singapore (2016). He also won the PASCAL VOC Challenge in Semantic Segmentation in 2010. His current research lives at the intersection of engineering of deep neural networks, theoretical machine learning and computational neuroscience. His goal is to develop empirically-grounded theories that explain deep neural networks and to build applications that have a positive societal impact.