ECE Seminar Series - Stratis Ioannidis

Title: Learning Saliency-Preserving Latent Representations

Abstract: We introduce an algorithm for learning salient feature representations through the explicit decomposition of salient and non-salient features into separate spaces. We show that our algorithm promotes learning low-dimensional, task-relevant features. We prove that the expected prediction deviation under input perturbations is upper-bounded by the dimension of the salient subspace and the so-called Hilbert-Schmidt Independence Criterion between inputs and representations. This establishes a link between robustness and latent representation compression in terms of the dimensionality and information preserved. Empirical evaluations on image classification tasks show that models trained with our algorithm primarily rely on salient input components, as indicated by reduced sensitivity to perturbations affecting non-salient features, such as image backgrounds.

Bio: Stratis Ioannidis is a professor in the Electrical and Computer Engineering department of Northeastern University, in Boston, MA, where he also holds a courtesy appointment with the Khoury College of Computer Sciences. He received his B.Sc. (2002) in Electrical and Computer Engineering from the National Technical University of Athens, Greece, and his M.Sc. (2004) and Ph.D. (2009) in Computer Science from the University of Toronto, Canada. Prior to joining Northeastern, he was a research scientist at the Technicolor research centers in Paris, France, and Palo Alto, CA, as well as at Yahoo Labs in Sunnyvale, CA. He is the recipient of an NSF CAREER Award, a Google Faculty Research Award, a Facebook Research Award, a Søren Buus Outstanding Research Award, a Martin W. Essigmann Outstanding Teaching Award, and several best paper awards. His research interests span machine learning, distributed systems, networking, optimization, and privacy.