Self-Supervised Learning with an Information Maximization Criterion

Authors: Serdar Ozsoy, Shadi Hamdan, Sercan Arik, Deniz Yuret, Alper Erdogan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments demonstrate that Cor Info Max achieves better or competitive performance results relative to the state-of-the-art SSL approaches. Section 5 provides the numerical experiments illustrating the performance of the proposed approach.
Researcher Affiliation Collaboration 1KUIS AI Center, Koc University, Turkey 2EEE Department, Koc University, Turkey 3CE Department, Koc University, Turkey 4Google Cloud AI Research, Sunnyvale, CA
Pseudocode Yes (See Appendix C for Cor Info Max pseudocode.)
Open Source Code Yes Cor Info Max s source code is publicly available in https://github.com/serdarozsoy/ corinfomax-ssl
Open Datasets Yes Datasets: We perform experiments on CIFAR-10, CIFAR-100 [40], Tiny Image Net [41], COCO [42], Image Net-100 and Image Net-1K [43] datasets1.
Dataset Splits Yes Then, we obtain the test accuracy results for the trained linear classifier based on the validation dataset.
Hardware Specification Yes For Image Net-100 and Image Net-1K, we pretrain our model on up to 8 A100 Cloud GPUs. The remaining datasets are trained using a single T4 and V100 Cloud GPU.
Software Dependencies No The paper mentions software components implicitly through the context of deep learning (e.g., training models), but does not provide specific version numbers for any libraries, frameworks, or languages used.
Experiment Setup Yes For pretraining, we use 1000 epochs with a batch size of 512 for CIFAR datasets, and 800 epochs with a batch size of 1024 for Tiny Image Net. [...] We use the SGD optimizer with a momentum of 0.9 and a weight decay of 1 10 4. The initial learning rate is 0.5 for CIFAR datasets and Tiny Image Net, 1.0 for Image Net-100, and 0.2 for Image Net-1K.