Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Adv-SSL: Adversarial Self-Supervised Representation Learning with Theoretical Guarantees
Authors: Chenguang Duan, Yuling Jiao, Huazhen Lin, Wensen Ma, Jerry Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our approach not only outperforms the existing methods across multiple benchmark datasets but is also supported by comprehensive end-to-end theoretical guarantees. [...] Through extensive experiments, we demonstrate that Adv-SSL significantly outperforms previous biased sample risk (Table 1), as well as several existing self-supervised learning approaches (Table 3). |
| Researcher Affiliation | Academia | 1School of Mathematics and Statistics, Wuhan University 2School of Artificial Intelligence, Wuhan University 3National Center for Applied Mathematics in Hubei, Wuhan University 4Hubei Key Laboratory of Computational Science, Wuhan University 5Center of Statistical Research, Southwestern University of Finance and Economics 6School of Statistics and Data Science, Southwestern University of Finance and Economics 7New Cornerstone Science Laboratory, Southwestern University of Finance and Economics 8Institute for Math & AI, Wuhan University EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Alternative Optimization Algorithm |
| Open Source Code | Yes | The Py Torch implementations can be found in supplementary material. [...] the implementation can be found in https://github.com/vincen-github/Adv-SSL. |
| Open Datasets | Yes | As the experiments conducted in existing self-supervised learning methods, we pretrain the representation on CIFAR-10, CIFAR-100 and Tiny Image Net, and subsequently conduct fine-tuning on each dataset with annotations. |
| Dataset Splits | Yes | As the experiments conducted in existing self-supervised learning methods, we pretrain the representation on CIFAR-10, CIFAR-100 and Tiny Image Net, and subsequently conduct fine-tuning on each dataset with annotations. [...] we train the linear classifier for 500 epochs using the Adam optimizer with corresponding labeled training set without data augmentation. |
| Hardware Specification | Yes | All experiments were conducted on a single Tesla V100 GPU. |
| Software Dependencies | No | The Py Torch implementations can be found in supplementary material. |
| Experiment Setup | Yes | We train for 1,000 epochs with a learning rate of 3 10 3 for CIFAR-10 and CIFAR-100, and 2 10 3 for Tiny Image Net. A learning rate warm-up is applied for the first 500 iterations of the optimizer, in addition to a 0.2 learning rate drop at 50 and 25 epochs before the training end. We use a mini-batch size of 256, and the dimension of the hidden layer in the projection head is set to 1024. The weight decay is set to 10 3. We adopt an embedding size (d ) of 512. The backbone network used in our implementation is Res Net-18. |