ML seminar: Efficient Optimization of a Convolutional Network with Gaussian Inputs

Speaker: Amir Globerson , Tel Aviv University

Date: Wednesday, March 01, 2017

Time: 5:00 PM to 6:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: 32-D463

Event Type:

Room Description:

Host: Tommi Jaakkola

Contact: Stefanie S. Jegelka, stefje@csail.mit.edu

Relevant URL:

Speaker URL: None

Speaker Photo:
None

Reminders to: seminars@csail.mit.edu, mitml@mit.edu

Reminder Subject: TALK: ML seminar: Amir Globerson: Efficient Optimization of a Convolutional Network with Gaussian Inputs

Abstract: Deep learning models are often successfully trained using
gradient descent, despite the worst case hardness of the underlying
non-convex optimization problem. The key question is then under what
conditions can one prove that optimization will succeed. Here we
provide, for the first time, a result of this kind for a one hidden
layer ConvNet with no overlap and ReLU activation. For this
architecture we show that learning is hard in the general case, but
that when the input distribution is Gaussian, gradient descent
converges to the global optimum in polynomial time. I will
additionally discuss an alternative approach to sidestepping the
complexity of deep learning optimization using improper learning.

Based on joint work with: Alon Brutzkus, Uri Heinemann, Roi Livni,
Elad Eban and Gal Elidan

Research Areas:

Impact Areas:

This event is not part of a series.

Created by Stefanie S. Jegelka at Monday, February 27, 2017 at 9:32 AM.