MetaMask: Revisiting Dimensional Confounder for Self-Supervised Learning

Authors: Jiangmeng Li, Wenwen Qiang, Yanan Zhang, Wenyi Mo, Changwen Zheng, Bing Su, Hui Xiong

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, our method achieves state-of-the-art performance on various benchmarks.1
Researcher Affiliation Academia Jiangmeng Li , Wenwen Qiang , Yanan Zhang University of Chinese Academy of Sciences Institute of Software Chinese Academy of Sciences Southern Marine Science and Engineering Guangdong Laboratory (Guangzhou) {jiangmeng2019, wenwen2018, yanan2018}@iscas.ac.cn Wenyi Mo Gaoling School of Artificial Intelligence Renmin University of China 2022101010@ruc.edu.cn Changwen Zheng Institute of Software Chinese Academy of Sciences Southern Marine Science and Engineering Guangdong Laboratory (Guangzhou) changwen@iscas.ac.cn Bing Su Gaoling School of Artificial Intelligence Renmin University of China Beijing Key Laboratory of Big Data Management and Analysis Methods subingats@gmail.com Hui Xiong Thrust of Artificial Intelligence The Hong Kong University of Science and Technology (Guangzhou) Guangzhou HKUST Fok Ying Tung Research Institute xionghui@ust.hk
Pseudocode Yes Algorithm 1 Meta Mask
Open Source Code Yes 1The implementation is available at https://github.com/lionellee9089/Meta Mask
Open Datasets Yes Model IN-200 [46] STL-10 [47] CIFAR-10 [46] CIFAR-100 [46]
Dataset Splits Yes Did you specify all the training details (e.g., data splits, hyper-parameters, how they were chosen)? [Yes] See Section 6, Appendix A.4, and the supplementary files.
Hardware Specification Yes Note that we adopt the official code of Barlow Twins and train on 8 GPUs of NVIDIA Tesla V100.
Software Dependencies No The paper mentions using "official code of Barlow Twins" but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For the comparisons demonstrated in Table 1, we uniformly set the batch size as 64, and we adopt a network with the 5 convolutional layers in Alex Net [48] as conv and a network with 2 additional fully connected layers as fc. [...] For the experiments in Table 2, the batch size is valued by 512, and Res Net-18 [51] is used as the backbone encoder. We adopt the data augmentation and other experimental settings following [2].