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..
Variational Supervised Contrastive Learning
Authors: Ziwen Wang, Jiajun Fan, Thao Nguyen, Heng Ji, Ge Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Trained exclusively on image data, our experiments on CIFAR-10, CIFAR-100, Image Net100, and Image Net-1K show that Var Con (1) achieves state-of-the-art performance for contrastive learning frameworks, reaching 79.36% Top-1 accuracy on Image Net1K and 78.29% on CIFAR-100 with a Res Net-50 encoder while converging in just 200 epochs; (2) yields substantially clearer decision boundaries and semantic organization in the embedding space, as evidenced by KNN classification, hierarchical clustering results, and transfer-learning assessments; and (3) demonstrates superior performance in few-shot learning than supervised baseline and superior robustness across various augmentation strategies. Our code is available at https://github.com/ziwenwang28/Var Contrast. |
| Researcher Affiliation | Academia | Ziwen Wang University of Illinois Urbana-Champaign EMAIL Jiajun Fan University of Illinois Urbana-Champaign EMAIL Thao Nguyen University of Illinois Urbana-Champaign EMAIL Heng Ji University of Illinois Urbana-Champaign EMAIL Ge Liu University of Illinois Urbana-Champaign EMAIL |
| Pseudocode | Yes | Figure 1: Var Con architectural flowchart and Pseudocode. Left: Input images are processed through an encoder network to produce ℓ2-normalized embeddings z. Class-level centroids wr are computed dynamically from mini-batch embeddings. The model determines sample s classification difficulty and applies confidence-adaptive temperature scaling τ2(z), which tightens constraints on challenging samples and relaxes them for well-classified examples. Right: Pseudocode implementation of our ELBO-derived loss function combining KL divergence and negative log-likelihood terms. |
| Open Source Code | Yes | Our code is available at https://github.com/ziwenwang28/Var Contrast. |
| Open Datasets | Yes | We evaluate our Var Con loss on four standard benchmarks: CIFAR-10, CIFAR-100 [40], Image Net100, and Image Net [52, 18], using the official test splits. |
| Dataset Splits | Yes | We evaluate our Var Con loss on four standard benchmarks: CIFAR-10, CIFAR-100 [40], Image Net100, and Image Net [52, 18], using the official test splits. |
| Hardware Specification | Yes | All experiments were conducted on 8 NVIDIA A100 GPUs. |
| Software Dependencies | No | The paper mentions "mixed-precision training (AMP)" but does not provide specific software names with version numbers for core libraries like PyTorch, TensorFlow, or CUDA. |
| Experiment Setup | Yes | Our best results are achieved with batch sizes of 512 for CIFAR-10/100 (200 epochs), 1,024 for Image Net-100 (200 epochs), and 4,096 for Image Net (350 epochs). Throughout, we use SGD with momentum 0.9 and weight decay 10 4 for smaller datasets including CIFAR-10, CIFAR-100, and Image Net-100 and LARS [63] optimizer for Image Net training to ensure stability. Full hyperparameter settings and optimization details are available in Appendix B.1. |