A mathematical model for automatic differentiation in machine learning
This work addresses a foundational problem in machine learning by providing a mathematical framework for automatic differentiation, which is incremental but clarifies relationships between program differentiation and nonsmooth functions.
The authors tackled the lack of a mathematical model for automatic differentiation in machine learning by developing a nonsmooth calculus and applying it to stochastic approximation methods, showing that usual methods avoid artificial critical points with probability one.
Automatic differentiation, as implemented today, does not have a simple mathematical model adapted to the needs of modern machine learning. In this work we articulate the relationships between differentiation of programs as implemented in practice and differentiation of nonsmooth functions. To this end we provide a simple class of functions, a nonsmooth calculus, and show how they apply to stochastic approximation methods. We also evidence the issue of artificial critical points created by algorithmic differentiation and show how usual methods avoid these points with probability one.