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..
Understanding Contrastive Learning via Gaussian Mixture Models
Authors: Parikshit Bansal, Ali Kavis, Sujay Sanghavi
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We corroborate our theory with experiments on CIFAR100; representations learned by Info NCE loss match the performance of LDA on clustering metrics. 6 Experiments We validate our theoretical findings with experiments on synthetic and real data. For the synthetic data, we study the effect of noise δ in Augmentation-enabled Distribution , rank r of the projection matrix and condition number of covariance matrix on learned representations. For the real data experiments, we compare self-supervised methods against baselines for clustering CIFAR100 dataset on low dimensional subspaces using K-Means. |
| Researcher Affiliation | Academia | Parikshit Bansal UT Austin EMAIL Ali Kavis UT Austin EMAIL Sujay Sanghavi UT Austin EMAIL |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. It primarily uses mathematical notation and textual descriptions for methods. |
| Open Source Code | No | 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We are planning to release our code after we prepare a clean and distributable version which is also de-anonymized. |
| Open Datasets | Yes | We corroborate our theory with experiments on CIFAR100; representations learned by Info NCE loss match the performance of LDA on clustering metrics. 6.3 Real-data experiments on CIFAR-100. 6.4 Real-data experiments on Image Net. |
| Dataset Splits | No | The paper describes sampling K=20 classes from CIFAR-100 and ImageNet for experiments and evaluating clustering performance using K-Means. It also mentions running methods for 5 different seeds corresponding to 5 random 20-class subsets. However, it does not explicitly provide details on how these sampled classes are further split into training, testing, or validation sets for learning the linear mappings or for evaluating the clustering algorithms in a traditional train/test split manner. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU models, or cloud instance types used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details such as library or solver names with version numbers (e.g., Python 3.8, PyTorch 1.9) needed to replicate the experiments. It mentions using 'gradient descent' and 'K-Means' but without specific software versions. |
| Experiment Setup | No | The paper mentions using 'gradient descent' for learning the mapping matrix A and describes data generation, sampling classes, and image preprocessing steps (subsampling, grayscaling, mean scaling). However, it does not provide specific details on critical hyperparameters such as learning rate, batch size, or the number of epochs used for training the models. |