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..
Enhancing Contrastive Learning with Variable Similarity
Authors: Haowen Cui, Shuo Chen, Jun Li, Jian Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the superiority of our approach, achieving gains of 2.1% on Image Net-100 and 1.4% on Image Net-1k compared with the state-of-the-art methods. Experimental results conducted on standard benchmarks demonstrate the superiority of our method, surpassing the state-of-the-art methods by 2.1% on Image Net-100 and 1.4% on Image Net-1k, respectively. In this subsection, we conduct ablation studies to evaluate the robustness of our method. |
| Researcher Affiliation | Academia | 1PCA Lab , Nanjing University of Science and Technology, China 2School of Intelligence Science and Technology, Nanjing University, China 3Center for Advanced Intelligence Project, RIKEN National Science Institute, Japan |
| Pseudocode | No | The paper describes mathematical formulations and loss functions but does not present a clearly labeled pseudocode or algorithm block. The methodology is explained in prose and equations. |
| Open Source Code | No | We will soon open-source the code, and the datasets used are publicly available for download by anyone. |
| Open Datasets | Yes | We pretrain the standard Res Net-50 [4] encoder on the Image Net-100 and Image Net-1k [45] datasets. We utilize the encoder pretrained on the Image Net-100 dataset and conduct linear evaluation on 7 datasets: CIFAR10/100 [47], Caltech101 [48], SUN397 [49], Food [50], Flowers [51], and Pets [52]. We conduct experiments on FC100 [55], CUB200 [56], and Plant [57] datasets. |
| Dataset Splits | Yes | The linear evaluation protocol on Image Net-100 and Image Net-1k datasets follows [12]. The linear evaluation protocol for transfer learning follows [35]. The training datasets are split into a train set and a validation set, with 90% for training and the remaining 10% for validation. |
| Hardware Specification | Yes | Specifically, we use 4 NVIDIA A100-SXM4-40GB GPUs to train these methods, where the batch size is set to 256. |
| Software Dependencies | No | The paper mentions optimizers like LARS [59], L-BFGS [60], and AdamW [65] but does not specify software dependencies (libraries, frameworks) with version numbers. |
| Experiment Setup | Yes | The encoder is pretrained for 200 epochs with the batch size of 256. Each model in φ is implemented with 2 fully connected layers, and each predictor in ψ consists of a single fully connected layer. We set ω = 0.5 following [35], λ = 0.5, and γ = 1 in the training loss, respectively. The learning rate is set to 0.03 and the temperature parameter for contrastive loss is 0.2. The projector consists of 2 MLP layers with an output dimension of 128. The memory queue size is 65536 and the exponential moving average (EMA) parameter is 0.999. The learning rate is set to 0.05. The projector and predictor consist of 3 MLP layers and 2 MLP layers with an output dimension of 2048, respectively. We follow the experimental settings in Mo Co-v3, including Adam W optimizer [65] with a linear learning rate warm-up for the first 40 epochs, a momentum of 0.9, and a weight decay of 0.1. A cosine learning rate schedule is applied to the encoder and predictor. The learning rate is set to 1.5e-4 and temperature is 0.2. |