Hadi is a postdoc at Princeton University. He previously worked at INRIA Paris as a postdoc researcher under the supervision of Professor Francis Bach. Hadi completed his Ph.D. in computer science in June 2020 in the Machine Learning Department of ETH Zurich under the supervision of Professor Thomas Hofmann. The focus of his research is optimization for (deep) neural networks.

SWS   Info

AG Audience

English

60 Minutes

Representation learning with neural networks automates feature extraction with less need for continental feature engineering; thereby achieving incredible performance in image, text, and strategy processing. However, the underlying mechanism of representation learning is not well understood. This limits applications of representation learning in critical tasks, such as cancer diagnoses and other medical decisions. In this talk, we propose a research plan for studying representation learning with three core research focuses: Random neural networks. By studying random neural networks, we shed light on the inner workings of the incredible performance of modern neural networks. We demonstrate how the study of random networks allows us to go beyond the conventional trial and error development of neural networks. Local optimality. Given a neural network, is it possible to improve its performance only by slight modifications of the network parameters? This is the focus of local optimization for representation learning. Our research highlights that local optimization requires more studies in modern representation learning with generative adversarial networks. Modeling. A mathematical study of learning dynamics is very challenging. Modeling facilitates the study of learning dynamics by omitting technical details of learning. For example, a continuous-time dynamical system may model an iterative learning method, bridging the gap between dynamical systems and representation learning.

Please contact MPI-SWS Office Team for link information

Susanne Girard

+49 631 9303 9605

--email hidden

passcode not visible

logged in users only