Revealing Distribution Discrepancy by Sampling Transfer in Unlabeled Data

Authors: Zhilin Zhao, Longbing Cao, Xuhui Fan, Wei-Shi Zheng

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experimental results This section presents a comparative analysis of I-Div 2 against existing methods for evaluating the distribution discrepancy between training and test samples. The detailed experimental setups are presented in Appendix D.1.
Researcher Affiliation Academia Zhilin Zhao1,2 Longbing Cao1 Xuhui Fan1 Wei-Shi Zheng2,3 1 School of Computing, Macquarie University, Australia 2 School of Computer Science and Engineering, Sun Yat-sen University, China 3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China
Pseudocode Yes Algorithm 1 Importance divergence
Open Source Code Yes The source code is publicly available at: https://github.com/Lawliet-zzl/I-div.
Open Datasets Yes We utilize two datasets, CIFAR10 [31] and SVHN [32], each comprising ten semantically unique classes.
Dataset Splits No The paper discusses training and test datasets but does not explicitly specify the use of a validation set or details about its split.
Hardware Specification Yes All experiments are conducted using the Py Torch framework on a single 64GB GPU.
Software Dependencies No The paper mentions using the "Py Torch framework" and other tools like "CLIP", "Res Net50", and "Vi T-B/16", but it does not provide specific version numbers for any of these software dependencies.
Experiment Setup Yes Unless otherwise specified, we set λ = 1 and γ = 1. ... Unless otherwise noted, our experiments use a standard sample size of M = 1000. We generate 100, 000 tuples to achieve significant distribution discrepancy in the positive pairs and minimal discrepancy in the negative pairs.