Free Lunch for Domain Adversarial Training: Environment Label Smoothing

Authors: YiFan Zhang, xue wang, Jian Liang, Zhang Zhang, Liang Wang, Rong Jin, Tieniu Tan

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate, both experimentally and theoretically, that ELS can improve training stability, local convergence, and robustness to noisy environment labels.
Researcher Affiliation Collaboration Yi-Fan Zhang1,2 , Xue Wang3, Jian Liang1,2, Zhang Zhang1,2, Liang Wang1,2, Rong Jin3 , Tieniu Tan1,2 1National Laboratory of Pattern Recognition (NLPR), Institute of Automation 2School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) 3 Machine Intelligence Technology, Alibaba Group.
Pseudocode No The paper does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes The code is avaliable at https://github.com/yfzhang114/Environment-Label-Smoothing.
Open Datasets Yes Rotated MNIST (Ghifary et al., 2015) consists of 70,000 digits in MNIST with different rotated angles where domain is determined by the degrees d {0,15,30,45,60,75}. PACS (Li et al., 2017b) includes 9, 991 images with 7 classes [...] from 4 domains. VLCS (Torralba & Efros, 2011) is composed of 10,729 images, 5 classes [...] from domains. Office-31 (Saenko et al., 2010) contains contains 4,110 images, 31 object categories in three domains. Office-Home (Venkateswara et al., 2017): consists of 15,500 images from 65 classes and 4 domains. ... Civil Comments-Wilds (Koh et al., 2021) contains 448,000 comments... Amazon-Wilds (Koh et al., 2021) contains 539,520 reviews... Rx Rx1-wilds (Koh et al., 2021) comprises images of cells... OGB-Mol PCBA (Koh et al., 2021) is a multi-label classification dataset... Spurious-Fourier (Gagnon-Audet et al., 2022) is a binary classification dataset... HHAR (Gagnon-Audet et al., 2022) is a 6 activities classification dataset...
Dataset Splits Yes The model selection that we use is test-domain validation, one of the three selection methods in (Gulrajani & Lopez-Paz, 2021). That is, we choose the model maximizing the accuracy on a validation set that follows the same distribution of the test domain. For DA tasks, all baselines implementation and hyper-parameters follows (Wang & Hou).
Hardware Specification Yes We conduct all the experiments on a machine with i7-8700K, 32G RAM, and four GTX2080ti.
Software Dependencies No The paper mentions software like 'Py Torch' and 'Distill BERT', but it does not provide specific version numbers for these software dependencies, which is required for reproducibility.
Experiment Setup Yes Table 9: Hyper-parameters for different benchmarks. Lrg,Decayg: learning rate and weight decay for the encoder and classifier; Lrd,Decayd: learning rate and weight decay for the domain discriminator; bsz: batch size during training; dsteps: the discriminator is trained dsteps times once the encoder and classifier are trained; Wreg: tradeoff weight for the gradient penalty; λ: tradeoff weight for the adversarial loss. The default β2 for Adam and Adam W optimizer is 0.99 and the momentum for SGD optimizer is 0.9.