arXiv (Cornell University) · 2021 · 16 citations · 29 references
ComputingComputer ScienceMathematicsArtificial IntelligenceMachine LearningOptimizationConvex OptimizationDeep LearningGradient-based OptimizationNeural NetworksOptimization For Machine LearningPartial Differential EquationsConic OptimizationConic ProgrammingGradient DescentImplicit DifferentiationImplicit Function TheoremNonsmooth OptimizationOperational CalculusOptimization Algorithms
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.
29
Learning representations by back-propagating errors
David E. Rumelhart, Geoffrey E. Hinton, Ronald J. Williams · Nature · 1986 · 30.9K citations
PyTorch: An Imperative Style, High-Performance Deep Learning Library
Adam Paszke, Sam Gross, Francisco Massa et al. · arXiv (Cornell University) · 2019 · 16.2K citations · Full text
Bradley Efron, Trevor Hastie, Iain M. Johnstone et al. · The Annals of Statistics · 2004 · 9.5K citations · Full text