Timezone: »
In view of training increasingly complex learning architectures, we establish a nonsmooth implicit function theorem with an operational calculus. Our result applies to most practical problems (i.e., definable problems) provided that a nonsmooth form of the classical invertibility condition is fulfilled. This approach allows for formal subdifferentiation: for instance, replacing derivatives by Clarke Jacobians in the usual differentiation formulas is fully justified for a wide class of nonsmooth problems. Moreover this calculus is entirely compatible with algorithmic differentiation (e.g., backpropagation). We provide several applications such as training deep equilibrium networks, training neural nets with conic optimization layers, or hyperparameter-tuning for nonsmooth Lasso-type models. To show the sharpness of our assumptions, we present numerical experiments showcasing the extremely pathological gradient dynamics one can encounter when applying implicit algorithmic differentiation without any hypothesis.
Author Information
Jérôme Bolte (Université Toulouse Capitole and TSE)
Tam Le (Université Toulouse 1 Capitole)
Edouard Pauwels (IRIT)
Tony Silveti-Falls (Toulouse School of Economics)
More from the Same Authors
-
2022 Poster: Automatic differentiation of nonsmooth iterative algorithms »
Jerome Bolte · Edouard Pauwels · Samuel Vaiter -
2021 Poster: Semialgebraic Representation of Monotone Deep Equilibrium Models and Applications to Certification »
Tong Chen · Jean Lasserre · Victor Magron · Edouard Pauwels -
2021 Poster: Numerical influence of ReLU’(0) on backpropagation »
David Bertoin · Jérôme Bolte · Sébastien Gerchinovitz · Edouard Pauwels -
2020 Poster: A mathematical model for automatic differentiation in machine learning »
Jérôme Bolte · Edouard Pauwels -
2020 Spotlight: A mathematical model for automatic differentiation in machine learning »
Jérôme Bolte · Edouard Pauwels -
2016 Poster: Sorting out typicality with the inverse moment matrix SOS polynomial »
Edouard Pauwels · Jean Lasserre