Theory Seminar: Automatic Differentiation of Tensor Expressions

Sören Laue (Uni Jena)
Automatic differentiation is a powerful tool that allows to compute derivatives not only of mathematical expressions but also of functions that are given as a computer program

Sören Laue (Uni Jena)
Herbert-Walther Lecture Hall G0.25
Wed 19. June 2019, 11:30 am

Abstract:

Automatic differentiation is a powerful tool that allows to compute derivatives not only of mathematical expressions but also of functions that are given as a computer program. While it has been used successfully in many areas it has recently gained considerable attention in the area of machine learning, especially in deep learning. In this talk, I will provide an introduction into the concept of automatic differentiation, explain the different algorithms for computing derivatives of computer programs, and provide examples. I will then focus on computing derivatives of problems involving not only scalars, but also vectors, matrices, and higher-order tensors. Finally, I will highlight the application of automatic differentiation to tensor networks.