Class-Distribution-Aware Pseudo-Labeling for Semi-Supervised Multi-Label Learning
Authors: Ming-Kun Xie, Jiahao Xiao, Hao-Zhe Liu, Gang Niu, Masashi Sugiyama, Sheng-Jun Huang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on multiple benchmark datasets confirm the efficacy of CAP in addressing the challenges of SSMLL problems. In this section, we first perform experiments to validate the effectiveness of the proposed method; then, we perform ablation studies to analyze the mechanism behind CAP. |
| Researcher Affiliation | Academia | 1Nanjing University of Aeronautics and Astronautics 2MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, Nanjing, China 3RIKEN Center for Advanced Intelligence Project 4The University of Tokyo, Tokyo, Japan |
| Pseudocode | No | The paper describes the method using mathematical formulations and descriptive text, but does not include a dedicated pseudocode or algorithm block. |
| Open Source Code | Yes | The implementation is available at https://github.com/milkxie/SSMLL-CAP. |
| Open Datasets | Yes | To evaluate the performance of the propose method, we conduct experiments on three benchmark image datasets, including Pascal VOC-2012 (VOC for short) 2 [12], MS-COCO-2014 (MS-COCO for short) 3 [25], and NUS-WIDE (NUS for short) 4 [7]. |
| Dataset Splits | Yes | For each dataset, we randomly sample a proportion p {0.05, 0.1, 0.15, 0.2} of examples with full labels while the others without any supervised information. MS-COCO is another widely used multi-label dataset, which consists of 82,081 training examples and 40,504 validation examples belonging to 80 different categories. In our experiments, the validation set is used for testing. |
| Hardware Specification | Yes | We perform all experiments on Ge Force RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions software components like ResNet-50, Rand Augment, Cutout, AdamW, and one-cycle policy scheduler, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | We employ Res Net-50 [16] pre-trained on Image Net [36] for training the classification model. We adopt Rand Augment [9] and Cutout [10] for data augmentation. We employ Adam W [29] optimizer and one-cycle policy scheduler [10] to train the model with maximal learning rate of 0.0001. The number of warm-up epochs is set as 12 for all datasets. The batch size is set as 32, 64, and 64 for VOC, MS-COCO, and NUS. Furthermore, we perform exponential moving average (EMA) for the model parameter θ with a decay of 0.9997. For all methods, we use the ASL loss as the base loss function, since it shows superiority to BCE loss [34]. We perform all experiments on Ge Force RTX 3090 GPUs. The random seed is set to 1 for all experiments. |