BIU learning club – Amir Globerson – Notions of simplicity in deep learning: From time series to images

Location:

Engineering building (1103), room 329

Title:

Notions of simplicity in deep learning: From time series to images 

Abstract:

It is standard practice indeep learning to train large models on relatively small datasets. This canpotentially lead to severe overfitting, but more often than not, test error isactually good. This phenomenon has prompted research on the so-called “ImplicitBias of Deep Learning Algorithms”. Here I will discuss our recent works onmultiple novel facets of this bias, and present theoretical and empiricalresults in different settings. In particular, I will discuss analysis ofimplicit bias in fine-tuning of large models, in learning temporal models(e.g., RNNs) and in labeling images with very few examples.

Short Bio:

Prof. Globerson received his BSc in computer science andphysics in 1997 from the Hebrew University, and his PhD in computationalneuroscience from the Hebrew University in 2006. After his PhD, he was apostdoctoral fellow at the University of Toronto and a Rothschild postdoctoralfellow at MIT. He joined the Hebrew University school of computer science in 2008,and moved to the Tel Aviv University School of Computer Science in 2015. He wasan associate editor for the Journal of Machine Learning Research, and theAssociate Editor in Chief for the IEEE Transactions on Pattern Analysis andMachine Intelligence. His work has received several paper awards (at NIPS,UAI, and ICML). He also serves as Research Scientist at Google in TelAviv. In 2018 he served as program co-chair for the UAI conference, and in 2019he was the general co-chair for UAI in Tel Aviv. In 2019 he receivedthe ERC consolidator grant.