Zoom Virtual Meeting:  https://cmu.zoom.us/j/92178295543?pwd=L2dwZU5SbDY5NzZZNzZ4ZmFUclRqQT09

Abstract:

Deep Neural Networks (DNNs) have revolutionized computer vision. We now have DNNs that achieve top results in many computer vision problems, including object recognition, facial expression analysis, and semantic segmentation, to name but a few. Unfortunately, the rise in performance has come with a cost.  DNNs have become so complex that are generally considered black-box models. We don’t know how a DNN learns, when does it start to  memorize and how well will it perform to unseen samples. In this talk, I will show how using Algebraic Topology can shed new light on some of these issues.  I will show that: 1). Functional topology of DNNs that learn is fundamentally different from those that simply memorise training data. 2). Topological descriptors of a trained DNN correlate with the generalization gap.  These findings can be used in several applications, including performing early stopping, detecting adversarial attacks and estimating the generalization gap without the need of a test set.

BIO:

Ciprian is currently Head of Research at Tawny GmbH, a Munich based startup developing
Emotional AI solutions for the consumer market. He has previously been Research Assistant at the University of Barcelona (Spain) and a fellow  of the Computer Vision Center, Barcelona.  Ciprian has received a PhD degree in 2019 from the University of Barcelona with a thesis on learning interpretable facial representations through deep neural networks. Ciprian’s research interests include deep learning, face and behavior analysis, affective computing, social signal processing, and human-computer interaction. His work has been published in top journals (TPAMI) and conferences (CVPR, ECCV). Most notably, his most recent paper about the topology of deep networks that generalise was a Best Paper Award Nominee in 2020 at CVPR.

Speaker Homepage: https://cipriancorneanu.github.io/

Sponsored in part by:   Facebook Reality Labs Pittsburgh