Discriminative Complementary-Label Learning with Weighted Loss
Authors: Yi Gao, Min-Ling Zhang
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate the performance of the proposed approach with comparative studies against state-of-the-art complementary-label learning approaches. We use L-UW and L-W to denote the proposed CLL approach instantiated with complementary loss function in Eq.(5) and Eq.(14) respectively. |
| Researcher Affiliation | Academia | 1School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China 2Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, China 3School of Computer Science and Engineering, Southeast University, Nanjing 210096, China. |
| Pseudocode | Yes | Algorithm 1 CLL with weighted loss |
| Open Source Code | Yes | The code is available at https://github.com/Yolkjustlike/complementary-label-learning. |
| Open Datasets | Yes | Following Ishida et al. (2017; 2019); Yu et al. (2018); Feng et al. (2020), three widely-used benchmark datasets, namely MNIST (Lecun et al., 1998), Fashion MNIST (Fashion) (Xiao et al., 2017), and Kuzushiji-MNIST (Kuzushiji) (Clanuwat et al., 2018), are used for experimental studies. |
| Dataset Splits | Yes | We divide the original training dataset into training and validation parts with proportion 9/1, where complementary labels are generated by randomly choosing one of the labels other than the ground-truth one (unbiased complementarylabel generation). |
| Hardware Specification | No | The paper states that experiments are implemented on “Colab” and mentions “The Big Data Center of Southeast University for providing the facility support on the numerical calculations in this paper”. However, it does not provide specific hardware details such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions “Py Torch (Paszke et al., 2019)” and “Adam (Kingma & Ba, 2015) optimization method”. While these are software components, specific version numbers for PyTorch or other libraries are not provided, which is required for reproducible software description. |
| Experiment Setup | Yes | Weight decay is set as 1e-4 and learning rate of 5e-5 is used for MNIST, Fashion and Kuzushiji. Adam (Kingma & Ba, 2015) optimization method is applied. For all datasets, the number of epoch and mini-batch size are set as 300 and 256 respectively. |