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..
Robust Label Proportions Learning
Authors: Jueyu Chen, Wantao Wen, Yeqiang Wang, Erliang Lin, Yemin Wang, Yuheng Jia
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments and ablation studies on multiple benchmarks demonstrate that RLPL achieves comparable state-of-the-art performance and effectively mitigates pseudo-label noise. |
| Researcher Affiliation | Academia | 1School of Artificial Intelligence, Southeast University, Nanjing, China 2Northwest A&F University, Xianyang, China 3School of Informatics, Xiamen University, Xiamen, China 4School of Computer Science and Engineering, Southeast University, Nanjing, China EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 RLPL: Robust Label Proportions Learning |
| Open Source Code | No | To protect privacy of our work, we will not open access the data and code until this paper is pubished. |
| Open Datasets | Yes | We utilized four standard benchmark datasets commonly employed in Learning from Label Proportions (LLP) research. These datasets are CIFAR-10, CIFAR-100 [14], SVHN [21], and Mini-Image Net [30]. |
| Dataset Splits | Yes | Both the CIFAR-10 and CIFAR-100 datasets [14] contain 50,000 training images and 10,000 test images. ... The SVHN dataset consists of 32 32 RGB images of digits, with 73,257 images for training and 26,032 for testing... Mini-Image Net... includes 100 classes, each with 80 images for training and 20 for testing... For each dataset, bags of a specified size M were formed by randomly selecting M samples from the training set, ensuring that samples in distinct bags do not overlap. |
| Hardware Specification | Yes | All experiments were conducted using a single NVIDIA RTX 4090 GPU with 24GB memory. |
| Software Dependencies | No | In the first stage, for training the encoder via unsupervised contrastive representation learning, we employ Sim CLR strategy. The encoder backbone is Res Net-18. The projection head in Sim CLR consists of a single linear layer. For Sim CLR training, we utilize the Adam optimizer, setting the learning rate to 1 10 3 and the weight decay to 1 10 6. Subsequently, the auxiliary classifier head, also a single MLP layer, is trained using the Adam optimizer with a learning rate of 1 10 4 and a weight decay of 1 10 7. |
| Experiment Setup | Yes | First Stage Training (Auxiliary Classifier) In the first stage, for training the encoder via unsupervised contrastive representation learning, we employ Sim CLR strategy. The encoder backbone is Res Net-18. The projection head in Sim CLR consists of a single linear layer. For Sim CLR training, we utilize the Adam optimizer, setting the learning rate to 1 10 3 and the weight decay to 1 10 6. Subsequently, the auxiliary classifier head, also a single MLP layer, is trained using the Adam optimizer with a learning rate of 1 10 4 and a weight decay of 1 10 7. ... Second Stage Training (Main Classifier) During the second stage for training the main classifier, the LLP-Proportion Penalty coefficient, denoted as λOT D, in the LLP-OTD module s cost function is set to 0.1. ... For the LLPMix training, the weight for the LLP-Consistency loss, denoted as w LLP in the combined objective function LLLP Mix, is set to 1 10 4. The main classifier is trained using the SGD optimizer. The initial learning rate is set to 0.02, with a momentum of 0.9 and a weight decay of 5 10 4. Backbone Architectures for Main Classifier ... For the CIFAR-10 and SVHN datasets, both our model and the baselines utilize WRN-28-2 as the backbone. For the CIFAR-100 dataset, WRN-28-8 is employed. On the Mini-Image Net dataset, Res Net-18 serves as the backbone. |