Solving large–scale non–smooth stochastic optimization problems

Machine learning focuses on writing computer programs that can ‘learn’ from data. This is becoming increasingly important, as we try to understand the huge amounts of complicated data we are collecting (both in academia and in industry). While machine learning is one of the key
tools we use analyze large datasets, it often has trouble dealing with the enormous sizes of modern datasets (for example, learning something about every webpage on the internet or about all users of Facebook). In a breakthrough paper in 2012, Francis Bach and Mark Schmidt
(along with another co–author) showed the surprising result that we can ‘learn’ from huge datasets at a similar speed to how we ‘learn’ from smaller datasets. However, there result assumed that our model was ‘differentiable’ and that our dataset had a finite size. In this internship, we will try to relax these assumptions. This will let us apply a much larger variety of machine learning techniques to huge datasets, and will further allow us to build systems that keep learning efficiently over time.

Faculty Supervisor:

Mark Schmidt

Student:

Partner:

École normale supérieure

Discipline:

Computer science

Sector:

University:

The University of British Columbia

Program:

Globalink Research Award