Information-Theoretic Analysis of Unsupervised Domain Adaptation

Authors: Ziqiao Wang, Yongyi Mao

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

Reproducibility Variable Result LLM Response
Research Type Experimental This paper uses information-theoretic tools to analyze the generalization error in unsupervised domain adaptation (UDA). We present novel upper bounds for two notions of generalization errors. [...] Specifically, we present two simple techniques for improving generalization in UDA and validate them experimentally. [...] Experiments are performed to verify the effectiveness of these strategies.
Researcher Affiliation Academia Ziqiao Wang & Yongyi Mao University of Ottawa {zwang286,ymao}@uottawa.ca
Pseudocode Yes Algorithm 1 Controlling Label Information (Appendix D.2, page 28)
Open Source Code No Appendix D states: 'Our code builds largely on the implementation from Gulrajani & Lopez-Paz (2021)3 and Nguyen et al. (2022)4.' Footnotes 3 and 4 provide GitHub links to these existing external implementations. The paper does not explicitly state that *their* specific modifications or code for this paper are released or available.
Open Datasets Yes Datasets We select two popular small datasets, Rotated MNIST and Digits, to compare the different methods. Rotated MNIST is built based on the MNIST dataset (Le Cun et al., 2010) [...] Digits consists of three sub-datasets, namely MNIST, USPS (Hull, 1994) and SVHN (Netzer et al., 2011)
Dataset Splits Yes Every dataset has a validation set, and the model selection scheme is based on the best performance achieved on the validation set of target domain during training (oracle).
Hardware Specification Yes We implemented our approach using Py Torch (Paszke et al., 2019) and conducted all experiments on NVIDIA Tesla V100 GPUs with 32 GB of memory.
Software Dependencies No The paper mentions 'Py Torch (Paszke et al., 2019)' which implies a version from the citation year, but it does not explicitly state a specific version number (e.g., PyTorch 1.9). Python is also mentioned without a version.
Experiment Setup Yes Other settings are also the same as Gulrajani & Lopez-Paz (2021) and Nguyen et al. (2022), for example, each algorithm is trained for 100 epochs. To select the hyperparameters (λ1 and λ2) for ERM-GP, ERM-KL, KL-GP and KL-CL, we perform random search. Specifically, λ1 is searched between [0.1, 0.9] and λ2 is searched between [10 6, 0.8].