Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
A Unified Domain Adaptation Framework with Distinctive Divergence Analysis
Authors: Zhiri YUAN, Xixu HU, Qi WU, Shumin MA, Cheuk Hang LEUNG, Xin Shen, Yiyan HUANG
TMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Based on the unified generalization bound, we propose new domain adaptation models that achieve transferability through domain-invariant representations and conduct experiments on real-world datasets that corroborate our theoretical findings. We believe these insights are helpful in guiding the future design of distribution-aligning domain adaptation algorithms. [...] In this section, we enrich the DA models by proposing new model variants that take divergence measures which are not covered in literature before, including total variation (TV), Neyman χ2, Pearson χ2, KL, Reverse KL, SH, JS and MMD with Laplacian kernel functions. [...] We test the efficacy of the new DA model variants and compare with a variety of baselines on three datasets: Office-Home, Digits and Office-31. [...] We record the accuracy results on the Office-Home tasks in Table 4, the results on the Digits tasks in Table 5, and the results on the Office-31 transfer tasks in Table 6. [...] In addition to the classification tasks, we also test the efficacy of the new DA model variants on two image regression datasets: MPI3D and d Sprites. |
| Researcher Affiliation | Academia | Zhiri Yuan EMAIL School of Data Science, City University of Hong Kong Xixu Hu EMAIL School of Data Science, City University of Hong Kong Qi Wu EMAIL School of Data Science, City University of Hong Kong Shumin Ma EMAIL Guangdong Provincial Key Laboratory of Interdisciplinary Research and Application for Data Science, BNU-HKBU United International College Cheuk Hang Leung EMAIL School of Data Science, City University of Hong Kong Xin Shen EMAIL Department of System Engineering and Engineering Management, The Chinese University of Hong Kong Yiyan Huang EMAIL School of Data Science, City University of Hong Kong |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. It provides mathematical proofs and descriptions of methodologies, but not in an algorithmic format. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing source code for the described methodology, nor does it include a link to a code repository. |
| Open Datasets | Yes | We test the efficacy of the new DA model variants and compare with a variety of baselines on three datasets: Office-Home, Digits and Office-31. All experiments in this section are run on Dell Precision 7920 with Intel Xeon Gold 6256 CPU at 3.6GHz, and a set of NVIDIA Quadro GV100 GPU cards. [...] For the Digits dataset, we investigate two digits datasets MNIST and USPS with two transfer tasks (M U and U M). [...] Office-31 is the most widely used dataset in DA literature. It consists of 4,652 images in 31 categories collected from three sources: Amazon (A), Webcam (W) and DSLR (D). [...] For the Office-Home and Office-31 experiments, Res Net-50 (He et al., 2016) pre-trained on the Image Net (Deng et al., 2009) is adopted as the backbone feature extractor. [...] In addition to the classification tasks, we also test the efficacy of the new DA model variants on two image regression datasets: MPI3D and d Sprites. |
| Dataset Splits | Yes | For the Digits dataset, we investigate two digits datasets MNIST and USPS with two transfer tasks (M U and U M). We train and evaluate all the methods following the splits and evaluation protocol from Long et al. (2018), which divides MNIST into 60,000 training images and 10,000 test images, and divides USPS into 7,291 training images and 2,007 test images, respectively. [...] We follow the standard experiment protocol for unsupervised DA from Ganin et al. (2016) and report the average accuracy for each experiment as in Acuna et al. (2021). |
| Hardware Specification | Yes | All experiments in this section are run on Dell Precision 7920 with Intel Xeon Gold 6256 CPU at 3.6GHz, and a set of NVIDIA Quadro GV100 GPU cards. |
| Software Dependencies | No | All algorithms are implemented in Py Torch. For models using f-divergences, we introduce a domain critic inspired by Shen et al. (2018) and Nowozin et al. (2016) to estimate the f-divergence in a min-max manner. The domain critic is a 3-layer neural network with Re LU as activation functions. |
| Experiment Setup | Yes | The classifiers are 2-layer neural networks with width 1024, 2048 and 500 correspondingly. We use separate neural networks to approximate the divergences. For the Office datasets, we adopt 3-layer neural networks with the same width as the classifiers. For the Digits dataset, we use a 2-layer neural network with width 500. [...] We take the mini-batch SGD optimizer with the Nestorov momentum 0.9 and the learning rates of the classifier are set to be 10 times to that of the feature extractor. [...] The regressors are a 2-layer convolutional neural network with Batch Norm and Re LU as the activation function, followed by an average pooling layer and a 1-layer neural network with width 1024. For models using f divergences, we introduce a domain critic inspired by (Nowozin et al., 2016) to estimate the f divergence in a min-max manner. The domain critic is a 2-layer convolutional neural network followed by a 1-layer neural network with Re LU as activation functions. We solve the min-max problem using gradient reversal layers motivated by (Ganin et al., 2016). We take the mini-batch SGD optimizer with the Nestorov momentum 0.95 and the learning rates of the regressors are set to be 10 times that of the feature extractor. |