Provable Domain Generalization via Invariant-Feature Subspace Recovery

Authors: Haoxiang Wang, Haozhe Si, Bo Li, Han Zhao

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, our ISRs can obtain superior performance compared with IRM on synthetic benchmarks. In addition, on three real-world image and text datasets, we show that both ISRs can be used as simple yet effective post-processing methods to improve the worst-case accuracy of (pre-)trained models against spurious correlations and group shifts. ... We conduct experiments on both synthetic and real datasets to examine our proposed algorithms.
Researcher Affiliation Academia Haoxiang Wang 1 Haozhe Si 1 Bo Li 1 Han Zhao 1 1University of Illinois at Urbana-Champaign, Urbana, IL, USA. Correspondence to: Haoxiang Wang <hwang264@illinois.edu>.
Pseudocode Yes Algorithm 1 ISR-Mean Algorithm 2 ISR-Cov
Open Source Code Yes The code is released at https: //github.com/Haoxiang-Wang/ISR.
Open Datasets Yes Waterbirds (Sagawa et al., 2019): This is a image dataset built from the CUB (Wah et al., 2011) and Places (Zhou et al., 2017) datasets. ... Celeb A (Liu et al., 2015): This is a celebrity face dataset... ... Multi NLI (Williams et al., 2017): This is a text dataset for natural language inference.
Dataset Splits Yes We choose the hyper-parameters that minimize the mean error over the validation split of all environments. ... early stop models at the epoch with the highest worst-group validation accuracy.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions using "Adam optimizer" and "logistic regression solver provided in scikit-learn", but does not specify version numbers for these software components or any other libraries.
Experiment Setup Yes for 10K full-batch Adam ... iterations. ... We choose the hyper-parameters that minimize the mean error over the validation split of all environments. ... early stop models at the epoch with the highest worst-group validation accuracy. ... Adopting the same hyperparameter as that of Table 1