Bayesian Uncertainty Matching for Unsupervised Domain Adaptation

Authors: Jun Wen, Nenggan Zheng, Junsong Yuan, Zhefeng Gong, Changyou Chen

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comparisons with state of the art unsupervised domain adaptation methods on three popular benchmark datasets demonstrate the superiority of our approach, especially on the effectiveness of alleviating negative transfer. (Abstract) and Extensive experimental results on standard domain-adaptation benchmarks demonstrate the effectiveness of the proposed method, outperforming current state-of-the-art approaches. (Introduction) and We compare our method with state-of-the-art domain-adaptation approaches on several benchmark datasets: USPS-MNIST-SVHN dataset [Hoffman et al., 2018], Office-31 dataset [Saenko et al., 2010], and the recently introduced Office-home dataset [Venkateswara et al., 2017]. (Section 4)
Researcher Affiliation Academia 1Qiushi Academy for Advanced Studies, Zhejiang University, Hangzhou, China 2College of Computer Science and Technology, Zhejiang University, Hangzhou, China 3Computer Science and Engineering Department, State University of New York at Buffalo 4Department of Neurobiology, Zhejiang University School of Medicine, Hangzhou, China
Pseudocode No The paper describes the proposed method using mathematical equations and descriptions in paragraph text, but it does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code No The paper does not include any statement about releasing source code for the described methodology or a link to a code repository.
Open Datasets Yes We compare our method with state-of-the-art domain-adaptation approaches on several benchmark datasets: USPS-MNIST-SVHN dataset [Hoffman et al., 2018], Office-31 dataset [Saenko et al., 2010], and the recently introduced Office-home dataset [Venkateswara et al., 2017].
Dataset Splits No The paper mentions using 'standard training sets' and 'reverse validation' for hyper-parameter selection, but it does not provide specific percentages or sample counts for training, validation, or test splits to reproduce the data partitioning.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions using Alex Net and adopting strategies from other works like DANN and JAN, but it does not provide specific version numbers for software dependencies or libraries (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes Hyper-parameters. To enable stable training, we progressively increase the importance of the adaptation loss and set λadv = 2 1+exp(γ m) 1, where γ = 10 and m denotes the training progress ranging from 0 to 1. We use a similar hyper-parameter selection strategy as in DANN, called reverse validation. We set λu = 0.25λadv to ensure uncertainty reduction. With τ = 1.5, we forward each sample T = 12 times to obtain prediction uncertainty. We set tu = 0.2, for adaptation loss re-weighting, and τc = 1.8 for source classification loss. We dropout all fully-connected layers with a dropout ration q = 0.5. Improvements are not observed with further dropout on convolution layers.