Reducing Divergence in Batch Normalization for Domain Adaptation
Authors: Ellen Yi-Ge, Mingjing Wu, Zhenghan Chen
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive empirical evaluation across diverse cross-domain benchmarks corroborates the superiority of RBN in augmenting inter-domain transferability. |
| Researcher Affiliation | Collaboration | 1Carnegie Mellon University 2Nanyang Technological University 3Microsoft (China) Co., Ltd |
| Pseudocode | No | The paper does not contain explicit pseudocode or algorithm blocks, instead describing methods through text and mathematical formulations. |
| Open Source Code | Yes | Code https://github.com/Ellen Yi Ge/RBN |
| Open Datasets | Yes | Datasets Office-31 (Saenko et al. 2010) is a seminal benchmark in the domain adaptation field. Image CLEF-DA 1 dataset, derived from the Image CLEF 2014 challenge, serves as another crucial benchmark for assessing domain adaptation methodologies. Office-Home (Li et al. 2019) represents a more expansive benchmark, comprising four diverse domains: Art (A), Clipart (C), Product (P), and Real World (R). Vis DA-2017 (Peng et al. 2017) presents a particularly challenging simulation-to-real scenario. 1http://imageclef.org/2014/adaptation |
| Dataset Splits | No | The paper mentions utilizing all labeled source data and all unlabeled target data, and describes Vis DA-2017 as structured across training, validation, and test domains. However, specific reproducible split percentages, sample counts, or citations to predefined standard splits for all datasets are not provided. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | Our implementation leverages the Pytorch framework, employing mini-batch stochastic gradient descent (SGD) for optimization. (No version numbers provided for PyTorch or other software.) |
| Experiment Setup | Yes | Our implementation leverages the Pytorch framework, employing mini-batch stochastic gradient descent (SGD) for optimization. We set the weight decay to 5 10 4, momentum to 0.9, and learning rate to 10 3. For feature extraction, we utilize Res Net-50 as the backbone architecture for Office-31, Image Cl EF-DA, and Office-Home datasets, while opting for a pre-trained Res Net-101 for the Vis DA-2017 dataset. Our RBN implementation involves substituting BN layers with Group Normalization (GN), incorporating RBNBlocks throughout the network architecture. Specifically, we replace a predetermined number of deeper layers. Our experimental protocol involves utilizing all labeled source data and all unlabeled target data, reporting the mean classification accuracy across five randomized experiments for each transfer task. All other training parameters remain consistent with the baseline setup. We systematically varied the batch size ({8, 16, 32, 64}) and the number of training epochs ({10, 20, 30, 40, 50, 60}). |