ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning
Authors: Mingyu Xu, Zheng Lian, Lei Feng, Bin Liu, Jianhua Tao
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on multiple benchmark datasets demonstrate that our method can achieve state-of-the-art performance on noisy PLL. |
| Researcher Affiliation | Academia | 1The State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences 2School of Artificial Intelligence, University of Chinese Academy of Sciences 3School of Computer Science and Engineering, Nanyang Technological University 4Department of Automation, Tsinghua University 5Beijing National Research Center for Information Science and Technology, Tsinghua University |
| Pseudocode | Yes | Algorithm 1: Pseudo-code of ALIM. Input: Dataset D with the estimated noise level η, predictive model f, warm-up epoch e0, number of epoch Emax, weighting coefficient λ, trade-off between losses λmix. Output: The optimized model f. |
| Open Source Code | Yes | Our code is available at: https://github.com/zeroQiaoba/ALIM. |
| Open Datasets | Yes | In the main experiments, we evaluate the performance on two benchmark datasets of noisy PLL, CIFAR-10 [23] and CIFAR-100 [23]. In Section 3.2, we also conduct experiments on fine-grained datasets (CUB-200 [17] and CIFAR-100H [24]) and consider severe noise. |
| Dataset Splits | Yes | Following the standard experimental setup in PLL [17, 25], we split a clean validation set from the training set to determine hyper-parameters. Then, we transform the validation set back to its original PLL form and incorporate it into the training set to accomplish model optimization. |
| Hardware Specification | Yes | All experiments are implemented with Py Torch [27] and carried out with NVIDIA Tesla V100 GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch [27]' but does not provide a specific version number for it or any other software dependencies. |
| Experiment Setup | Yes | There are mainly three user-specific parameters in ALIM: λ, λmix, and e0. Among them, λ controls the trade-off between the initial candidate set and model outputs. This paper proposes two selection strategies for λ, i.e., manually and adaptively adjusted strategies. For the first one, we treat λ as a hyper-parameter and select it from {0.1, 0.2, 0.3, 0.4, 0.5, 0.7}. For the second one, we automatically determine λ using the estimated noise level. λmix controls the trade-off between the PLL loss and the mixup loss, and we set λmix = 1.0 as the default parameter. e0 is the start epoch of ALIM, and we select it from {20, 40, 80, 100, 140}. Following the standard experimental setup in PLL [17, 25], we split a clean validation set from the training set to determine hyper-parameters. Then, we transform the validation set back to its original PLL form and incorporate it into the training set to accomplish model optimization. To optimize all trainable parameters, we choose the SGD optimizer with a momentum of 0.9 and set the weight decay to 0.001. We set the initial learning rate to 0.01 and adjust it using the cosine scheduler. |