Nonsmooth Implicit Differentiation: Deterministic and Stochastic Convergence Rates

Authors: Riccardo Grazzi, Massimiliano Pontil, Saverio Salzo

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also present illustrative experiments confirming our analysis.
Researcher Affiliation Academia 1CSML, Istituto Italiano di Tecnologia, Genoa, Italy 2Department of Computer Science, University College, London, UK 3Dipartimento di Ingegneria Informatica, Automatica e Gestionale, Universit a La Sapienza, Rome, Italy.
Pseudocode Yes Algorithm 1 NSID; Algorithm 2 NSID-Bilevel; Algorithm 3 Stochastic fixed point iterations
Open Source Code Yes We provide the code to reproduce our experiments at https://github.com/prolearner/ nonsmooth_implicit_diff
Open Datasets Yes For the data poisoning setup we use the MNIST dataset.
Dataset Splits Yes We split the MNIST original train set into 30K example for training and 30K examples for validation.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory, or cloud instance types used for the experiments.
Software Dependencies No We implement NSID by relying on Py Torch automatic differentiation for the computation of Jacobian-vector products. For AID and ITD, we use the existing Py Torch implementations. (No version numbers provided for PyTorch).
Experiment Setup Yes The appropriate choice for the step-size guarantees that Φ is a contraction, in our case we set it equal to 2{p L µ 2λ2q... We set the regularization parameters λ p0.02, 0.1q... we set it [minibatch size] equal to 10% of the training set... we set ηi a1{pa2 iq for (N)SID dec and η a1{a2 for (N)SID const, where a1 b1β and a2 b2β... In particular we set a1 0.5, a2 2 for the synthetic Elastic net experiment and a1 2, a2 0.01 for Data poisoning.