Towards Calibrated Deep Clustering Network
Authors: Yuheng Jia, Jianhong Cheng, Hui LIU, Junhui Hou
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the proposed calibrated deep clustering model not only surpasses the state-of-the-art deep clustering methods by 5 on average in terms of expected calibration error, but also significantly outperforms them in terms of clustering accuracy. |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, Southeast University 2Yam Pak Charitable Foundation School of Computing and Information Sciences, Saint Francis University 3Department of Computer Science, City University of Hong Kong EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1: The training process of Calibrated Deep Clustering |
| Open Source Code | Yes | Code is available at https://github.com/Cheng Jian H/CDC. |
| Open Datasets | Yes | We conducted experiments on six widely used benchmark datasets, CIFAR-10, CIFAR-20 (Krizhevsky et al., 2009), STL-10 (Coates et al., 2011), Image Net-10, Image Net-Dogs (Chang et al., 2017), and Tiny-Image Net (Le & Yang, 2015). |
| Dataset Splits | Yes | Supervised Baseline. We trained on the training datasets of CIFAR-10, CIFAR-20, STL10 and Tiny-Image Net, and evaluated the results on the test or validation datasets. Due to a lack of the designated test datasets on Image Net-10 and Image Net-Dogs, we reported the results on the training datasets. |
| Hardware Specification | Yes | Moreover, All experiments are conducted on an NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | For CDC, we trained the model for 100 epochs using the Adam optimizer with a learning rate of 5e-5 for the encoder and 1e-4 for the MLP. We speed up K-Means with K-Means++ initialization and the Py Torch implementation as in Pro Pos (Huang et al., 2022). The paper does not provide specific version numbers for software. |
| Experiment Setup | Yes | For CDC, we trained the model for 100 epochs using the Adam optimizer with a learning rate of 5e-5 for the encoder and 1e-4 for the MLP. The learning rate of the encoder on CIFAR-20 and Tiny-Image Net was adjusted to 1e-5 for better learning of noisy pseudo labels. We set B=1,000 for CIFAR-10, CIFAR-20, and STL-10, B=500 for Image Net-10 and Image Net-Dogs, B=5,000 for Tiny-Image Net. We set K=500 for CIFAR-10, K=40 for CIFAR-20 and Image Net-Dogs, K=150 for STL-10 and Image Net-10, and K=1,000 for Tiny-Image Net. |