Federated Generalized Novel Category Discovery with Prompts Tuning
Authors: Lei Shen, Nan Pu, Zhun Zhong, Mingming Gong, Dianhai Yu, Chengqi Zhang, Bo Han
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments on both generic and fine-grained datasets like CIFAR-100 and CUB-200, and show that our method is comparable to the FL version of sim GCD and surpasses other baselines with significantly fewer parameters to transmit. |
| Researcher Affiliation | Collaboration | Lei Shen EMAIL TMLR Group, Department of Computer Science Hong Kong Baptist University Nan Pu EMAIL Department of Information Engineering and Computer Science University of Trento Zhun Zhong EMAIL School of Computer and Information Hefei University of Technology Mingming Gong EMAIL School of Mathematics and Statistics University of Melbourne Dianhai Yu EMAIL Baidu Inc. Chengqi Zhang EMAIL Department of Data Science and Artificial Intelligence Hong Kong Polytechnic University Bo Han EMAIL TMLR Group, Department of Computer Science Hong Kong Baptist University |
| Pseudocode | No | The paper describes the methodology and framework using text, mathematical formulations, and a diagram (Figure 2), but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing code, links to code repositories, or mentions of code being available in supplementary materials for the described methodology. |
| Open Datasets | Yes | To fairly evaluate the performance of Fed GCD methods, we conduct a comparison for all methods on two commonly-used generic image classification datasets (i.e., CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009)) and two fine-grained image classification datasets (i.e., CUB-200 (Welinder et al., 2010), Oxford-Pet (Parkhi et al., 2012). |
| Dataset Splits | Yes | We leverage the Latent Dirichlet Sampling (Dir) partition method (Hsu et al., 2020) to split the training set into N L subsets, each of which is stored in local clients as its local training data. After partitioning the whole training dataset into heterogeneous subsets, we sample a subset of half the classes as Seen categories in the original training set, and only local data from these classes are treated as potentially labeled data. In these potentially labeled data, 50% of instances of each labeled class are randomly sampled to form the labeled set. The remaining local data are treated as unlabeled data when training. We set N L = 5 and partition the whole dataset into local data with Dir(0.05) in our main experiments and explore more settings in Section 5.5 and Section 5.4. |
| Hardware Specification | Yes | We conduct experiments using NVIDIA V100 32GB GPU, 20 Intel(R) Xeon(R) Silver 4114 CPUs @ 2.20GHz. |
| Software Dependencies | Yes | The operating system is Oracle Linux 8 (x86_64) UEK Release 6. The pytorch version is 1.12.1. The numpy version is 2.0.1. The cuda version is 11.8. |
| Experiment Setup | Yes | We set the batch size to 128 for all methods and datasets. All methods are optimized by SGD with a momentum of 0.9 and weight decay of 1 10 4 for 200 epochs with a cosine annealing schedule starting from a learning rate of 0.1. All methods use dino-vitb16 (Caron et al., 2021) as the backbone following the previous setting in GCD works (Vaze et al., 2022; Pu et al., 2023a). The hyper-parameters λ controlling weight of supervised and self-supervised learning are set to 0.35 for Fed-GCD, Fed-Sim GCD, and Fed GCD-P. During training process, we define the initial random seed as 2023, then the random seed is increased by 1 every global round, while when federatedly partitioning the datasets, the random seed is set as 0. Our local training epoch is set as 1. |