BlueJeans Meeting ID: 919103457
Join Here
Abstract
Algorithmic or automatic, differentiation (AD or autodiff) is a technique for transforming algorithms that compute some mathematical function into algorithms that compute the derivatives of that function. AD techniques combine rules for differentiating the functions intrinsic to a given programming language with strategies for applying the chain rule. AD has been used extensively for computing derivatives that are used for sensitivity analysis, optimization, parameter (state) estimation etc. Derivatives, mostly in the form of gradients and Hessians, are ubiquitous in machine learning as well.