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..
Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning
Authors: Achleshwar Luthra, Tianbao Yang, Tomer Galanti
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experiment with the following datasets CIFAR10 and CIFAR100 [85], mini Image Net [86], Tiny-Image Net [87], SVHN [88] and Image Net-1K [89]. In several experiments we monitor LDCL (see (1)) and LNSCL (see 3.1). To evaluate the quality of learned representations, we use two methods: the Nearest Class-Center Classification (NCCC) accuracy [83], and linear probing accuracy [95, 96, 97]. |
| Researcher Affiliation | Academia | EMAIL Department of Computer Science and Engineering Texas A&M University |
| Pseudocode | No | The paper describes theoretical results and experimental setups, but it does not contain any clearly labeled pseudocode or algorithm blocks. Procedures are described in narrative text. |
| Open Source Code | Yes | The code and project page of the paper are available at [code, project page]. |
| Open Datasets | Yes | Datasets. We experiment with the following datasets CIFAR10 and CIFAR100 [85], mini Image Net [86], Tiny-Image Net [87], SVHN [88] and Image Net-1K [89]. |
| Dataset Splits | Yes | CIFAR10 and CIFAR100 both consist of 50000 training images and 10000 validation images with 10 classes and 100 classes, respectively, uniformly distributed across the dataset, i.e., CIFAR10 has 5000 samples per class and CIFAR100 has 500 samples per class. mini-Image Net also has 5000 test images on top of 50000 train and 10000 validation images, with 100 of 1000 classes from Image Net-1K [89] (at the original resolution). Tiny-Image Net contains 100000 images downsampled to 64 64, with total 200 classes from IM-1K. Each class has 500 training, 50 validation, and 50 test images. SVHN consists of digit classification data with 10 classes from real-world images, organized into "train" (73,257 samples), "test" (26,032 samples), and "extra" (531,131 samples) splits. For scalability verification, we combine the "train" and "extra" splits during training for validating Thm. 1 (shown in Fig. 3). Image Net-1K spans 1000 object classes and contains 1,281,167 training images, 50,000 validation images and 100,000 test images. We use "train" split for training and "validation" split for reporting test results. |
| Hardware Specification | Yes | All models were trained on a single node with two 94 GB NVIDIA H100 GPUs. |
| Software Dependencies | No | The paper mentions "Our models are implemented in Py Torch [101]" but does not provide specific version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | We trained our models with the Sim CLR [5] algorithm. We use a Res Net-50 [90] encoder with a width-multiplier factor of 2 and Vi T-Base [91]. Both are followed by a projection head with a standard two-layer MLP architecture composed of: Linear(2048 2048) Re LU Linear(2048 128). In order to minimize the loss, we adopt the LARS optimizer [92]... For LARS, we set the momentum to 0.9 and the weight decay to 1e 6. All experiments are carried out with a batch size of B = 1024. The base learning rate is scaled with batch size as 0.3 B/256 , following standard practice [5]. We employ a warm-up phase [93] for the first 10 epochs, followed by a cosine learning rate schedule without restarts [94] for the remaining epochs. |