Date: Friday, February 23
Time: 11:00am – 12:00pm
Location: ENGR 4201
Zoom Link: https://gmu.zoom.us/j/94538584246?pwd=WndHRHVMZ3N6OHBWcVlmaUg2QlI3UT09
Meeting ID: 945 3858 4246
Passcode: 329439
Title:
Two Losses Are Better Than One: Faster Optimization Using a Cheaper Proxy
Abstract:
In many machine learning problems, it is difficult to directly access the training objective making it difficult to apply popular training algorithms like stochastic gradient descent, which requires computing the gradient of the loss at each
step. For example, it may be difficult or costly to acquire training samples from the data distribution, making it expensive to estimate the gradient of the loss using samples; or, when training a robot to complete some task, computing the gradient of the
loss could require waiting for this robot to physically attempt the task, which may be time consuming and may risk damage to the robot or its surroundings.
In this talk, I will describe a novel approach to optimizing such hard-to-access training objectives by leveraging a "proxy loss"---any alternative objective whose gradients are easier to compute. I will discuss convergence guarantees for our
algorithm on convex functions, which demonstrate that the number of gradient computations needed from the expensive target objective can be greatly reduced when the proxy is a sufficiently good approximation of the target. I will also describe how such proxy
losses might be obtained in several realistic machine learning settings, for instance, using synthetic data, robotic simulators, and more.
Bio:
Blake Woodworth is an assistant professor at The George Washington University. His research focuses on the intersection between optimization theory and machine learning. Prior to joining GWU this year, Blake was a postdoctoral researcher at Inria
in Paris under the supervision of Francis Bach. He received his PhD in 2021 from the Toyota Technological Institute at Chicago advised by Nati Srebro.