On Balancing Bias and Variance in Unsupervised Multi-Source-Free Domain Adaptation
Authors: Maohao Shen, Yuheng Bu, Gregory W. Wornell
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on multiple datasets validate our theoretical analysis and demonstrate the state-of-art performance of the proposed algorithm, especially on some of the most challenging datasets, including Office-Home and Domain Net. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, USA 2Department of Electrical & Computer Engineering, University of Florida, Gainesville, USA. |
| Pseudocode | Yes | Algorithm 1 Selective Self-training for MSFDA |
| Open Source Code | No | The paper does not contain any explicit statement about making the source code available or provide a link to a code repository. |
| Open Datasets | Yes | Datasets We conduct extensive evaluations of our methods using the following four benchmark datasets: Digits Five (Peng et al., 2019) contains five different domains, including MNIST (MN), SVHN (SV), USPS (US), MNISTM (MM), and Synthetic Digits (SY). Office-31 (Saenko et al., 2010) contains 31 categories collected from three different office environments, including Amazon(A), Webcam (W), and DSLR (D). Office-Home (Venkateswara et al., 2017) is a more challenging dataset with 65 categories collected from four different office environments, including Art (A), Clipart (C), Real-world (R), and Product (P). Domain Net (Peng et al., 2019) is so far the largest and most challenging domain adaptation benchmark, which contains about 0.6 million images with 345 categories collected from six different domains, including Clipart (C), Infograph (I), Painting (P), Quickdraw (Q), Real (R), and Sketch (S). |
| Dataset Splits | No | The paper refers to using 'training data' and 'target domain data' but does not explicitly provide details about a specific validation dataset split, percentages, or methodology for creating such a split. |
| Hardware Specification | Yes | All experiments are implemented in Py Torch using Tesla V100 GPUs with 32 GB memory. |
| Software Dependencies | No | The paper mentions 'Py Torch' but does not provide specific version numbers for it or any other software libraries used. |
| Experiment Setup | Yes | For model optimization, we use SGD optimizer with momentum value 0.9 and weight decay 10 3, the learning rate for backbone, bottleneck layer, and classifier layer is set to 10 2, 10 2 and 10 3, respectively. For domain aggregation weights optimization, we also use SGD optimizer with learning rate 10 1 without weight decay. The maximum number of training iterations is set to 20. The other hyper-parameters are summarized in Table 9, where bs denotes the batch size, and itr denotes the training iteration. |