Department of Computer Science
Johns Hopkins University
Meeting ID: 937 3284 8196
Abstract: Dropout is a popular algorithmicregularization technique for training deep neural networks. While it has beenshown effective across a wide range of machine learning tasks ? like many otherpopular heuristics in deep learning ? dropout lacks a strong theoreticaljustification. In this talk, we present statistical and computational learningtheoretic guarantees for dropout training in several machinelearning models,including matrix sensing, deep linear networks, and two-layer ReLUnetworks. This talk is based primarily on the following two papers:https://arxiv.org/pdf/2003.03397.pdf, https://arxiv.org/pdf/2010.12711.pdf.
Bio: Poorya Mianjy is a Ph.D. candidatein the Department of Computer Science at the Johns Hopkins University, advisedby Raman Arora. He is interested in theoretical machine learning, and inparticular, the theory of deep learning.