Gradient Matching for Domain Generalization

Authors: Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, Gabriel Synnaeve

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments on the WILDS benchmark, which captures distribution shift in the real world, as well as the DOMAINBED benchmark that focuses more on syntheticto-real transfer. Our method produces competitive results on both benchmarks, demonstrating its effectiveness across a wide range of domain generalization tasks.
Researcher Affiliation Collaboration Yuge Shi University of Oxford yshi@robots.ox.ac.uk Jeffrey Seely Meta Reality Labs jseely@fb.com Philip H.S. Torr University of Oxford philip.torr@eng.ox.ac.uk N. Siddharth The University of Edinburgh & The Alan Turing Institute n.siddharth@ed.ac.uk Awni Hannun Facebook AI Research awni@fb.com Nicolas Usunier Facebook AI Research usunier@fb.com Gabriel Synnaeve Facebook AI Research gab@fb.com
Pseudocode Yes Algorithm 1 Fish. Algorithm 2 Direct optimization of IDGM. Algorithm 3 Black fonts denote steps used in both algorithms, colored fonts are steps unique to Fish or Reptile. Algorithm 4 Smoothed version of Fish, which allows to get approximate gradients for the general form of Equation (4). Algorithm 5 Function GIP. Algorithm 6 Algorithm of collecting gradient inner product g for Fish and ERM both before and after updates.
Open Source Code Yes Code is available at https://github.com/YugeTen/fish.
Open Datasets Yes We perform experiments on the WILDS benchmark (Koh et al., 2020)... as well as the DOMAINBED benchmark (Gulrajani and Lopez-Paz, 2020).
Dataset Splits Yes Table 5: Details of the 6 WILDS datasets we experimented on. Dataset... # Examples train val test... # Domains train val test
Hardware Specification Yes Our experiments can be replicated with 1500 GPU hours on NVIDIA V100.
Software Dependencies No The paper mentions optimizers (SGD, Adam) and model architectures (ResNet, DenseNet, DistilBERT) but does not provide specific version numbers for these or other software libraries or dependencies.
Experiment Setup Yes Table 13: Hyperparameters for ERM. We follow the hyperparameters used in WILDS benchmark. Table 14: Hyperparameters for Fish.