Скачать или смотреть Hongyang Zhang - Understanding and improving generalization in multitask and transfer learning

Presentation given by Hongyang Zhang on May 4th 2022 in the one world seminar on the mathematics of machine learning on the topic "Understanding and improving generalization in multitask and transfer learning".

Abstract: A broad research question in ML is how to build models that generalize across various tasks. This talk will describe my recent works addressing this question. First, I will talk about some new analyses of multitask learning, which explains negative transfer using recent developments in deep learning theory. Second, I will talk about transfer learning from pretrained deep neural networks. I will describe our approach to avoid overfitting, with provable generalization using Hessians. Lastly, I will briefly talk about a related multigroup learning problem, and a recent contrastive learning approach for solving this problem.