University of Cambridge > > Cambridge University Physics Society > Energy, entropy and the physics of deep learning

Energy, entropy and the physics of deep learning

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Akshat Pandey.

Deep learning has achieved beyond-human accuracy in a plethora of challenging tasks, ranging from image recognition to gameplay. Nonetheless, why deep learning “works” is thus far an open question. In my talk, I will argue that physical concepts such as energy and entropy allow us to explain the surprising efficacy of deep learning. I will show that stochastic gradient descent, a machine learning algorithm that is commonly used, can be mapped to the familiar physics of Brownian motion albeit with a spatially anisotropic noise that is crucial to its success. Moreover, theoretical tools used to analyse spin glasses and energy landscapes give us important insights about the structure of loss functions that are typically encountered in deep learning.

This talk is part of the Cambridge University Physics Society series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2024, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity