Exploring a Principled Framework for Deep Subspace Clustering
Authors: Xianghan Meng, Zhiyuan Huang, Wei He, Xianbiao Qi, Rong Xiao, Chun-Guang Li
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on the synthetic data and six benchmark datasets to verify our theoretical findings and demonstrate the superior performance of our proposed deep subspace clustering approach. |
| Researcher Affiliation | Collaboration | Xianghan Meng , Zhiyuan Huang & Wei He Beijing University of Posts and Telecommunications, Beijing 100876, P.R. China EMAIL Xianbiao Qi & Rong Xiao Intellifusion, Shenzhen, P.R. China Chun-Guang Li Beijing University of Posts and Telecommunications, Beijing 100876, P.R. China EMAIL |
| Pseudocode | Yes | Algorithm 1 Scalable & Efficient Implementation of PRO-DSC via Differential Programming |
| Open Source Code | Yes | To ensure the reproducibility of our work, we have released the source code. |
| Open Datasets | Yes | All datasets used in our experiments are publicly available, and we have provided a comprehensive description of the data processing steps in Appendix B.1. |
| Dataset Splits | Yes | For all the datasets except for Image Net-Dogs, we train the network to implement PRO-DSC on the train set and test it on the test set to validate the generalization of the learned model. For Image Net-Dogs dataset which does not have a test set, we train the network to implement PRO-DSC on the train set and report the clustering performance on the training set. For a direct comparison, we conclude the basic information of these datasets in Table B.1. |
| Hardware Specification | Yes | all the experiments are conducted on a single NVIDIA RTX 3090 GPU and Intel Xeon Platinum 8255C CPU. |
| Software Dependencies | No | The paper mentions software like "SGD optimizer" and "scikit-learn" but does not specify their version numbers. |
| Experiment Setup | Yes | We train the network by the SGD optimizer with the learning rate set to η = 10 4, and the weight decay parameters of f( ; Ψ) and h( ; Ψ) are set to 10 4 and 5 10 3, respectively. We set α = d 0.1 nb for all the experiments. We summarize the hyper-parameters for training the network to implement PRO-DSC in Table B.2. |