Anders Karlsson, Geneva/Uppsala: Metric spaces, an ergodic theorem and theoretical aspects of deep learning
Overview
- Date:Starts 9 May 2023, 15:30Ends 9 May 2023, 16:30
- Location:Euler, Skeppsgränd 3
- Language:English
Abstract: Deep neural networks, which just means the compositions of certain types of non-linear mathematical functions, have been instrumental for the recent rise of artificial intelligence (AI). Their success has been a great surprise, but it is often noted that to a large extent nobody understands why it works so well, a theory is missing. This is related to some of the limitations and problems with AI, such as reliability, biases and computational complexity, that currently are much discussed. Benny Avelin and I displayed semi-invariant distances for some of the more popular choices of layer functions. Moreover, in several ways deep learning gives rise to the composition (sometimes more than hundreds of) of randomly selected transformations, and then there is a general noncommutative ergodic theorem that applies and guarantees a certain regularity that recent research suggests may be advantageous for learning.
Fika 15.00-15.25 in the common room.
