Bayesian Domain Adaptation with Gaussian Mixture Domain-Indexing

Authors: Yanfang Ling, Jiyong Li, Lingbo Li, Shangsong Liang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We verify the effectiveness of GMDI via experimental comparison and analysis. In particular, we answer three research questions: (RQ1) Can the performance of GMDI for domain adaptation outperform baselines? (RQ2) How effective is the global domain indices inferred by GMDI? (RQ3) How does the number of mixture components K affect results? Additional experimental results are available in Appendix K.
Researcher Affiliation Collaboration Yanfang Ling Sun Yat-sen University lingyf3@mail2.sysu.edu.cn Jiyong Li Sun Yat-sen University lijy373@mail2.sysu.edu.cn Lingbo Li Inf Mind Technology Ltd lingbo@infmind.ai Shangsong Liang Sun Yat-sen University liangshangsong@gmail.com
Pseudocode Yes The procedure of our proposed model GMDI is summarized by the pesuedo codes in Algorithm 1. Let φ represent the parameters of the distribution p( ) in the generative process. Algorithm 1 Bayesian Domain Adaptation with Gaussian Mixture Domain-Indexing.
Open Source Code Yes Source code is publicly available from https://github.com/lingyf3/GMDI.
Open Datasets Yes Datasets. We compare GMDI with existing DA methods on the following datasets (see Appendix H and Appendix I for more details): Circle [36] is used for binary classification task. DG-15 and DG60 [40] are synthetic datasets used for binary classification task. TPT-48 [40] dataset is a real-world dataset used for regression task. Comp Cars [43] dataset is a real-world dataset for 4-way classification task.
Dataset Splits No The paper defines source and target domains for training and testing, but it does not specify explicit numerical splits for training, validation, and testing sets (e.g., 80/10/10 percentages or exact sample counts for each split).
Hardware Specification Yes We run experiments on a single machine using 1 NVIDIA Ge Force RTX 2080Ti with 11GB memory, 56 Intel Xeon CPUs (E5-2680 v4 @ 2.40GHz).
Software Dependencies No The paper mentions implementing the model based on VDI's code but does not specify software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow versions).
Experiment Setup Yes We set the maximum number of mixture components K from 2,3, and the concentration parameter α to 1 throughout the experiments. Except for DG-15 and DG-60 datasets, which have a batch size of 32, all other datasets use a batch size of 16. Our model is trained with 20 to 100 warmup steps, learning rates ranging from 1 10 5 to 1 10 3, and λ ranging from 0.1 to 1.