Towards Distribution-Agnostic Generalized Category Discovery

Authors: Jianhong Bai, Zuozhu Liu, Hualiang Wang, Ruizhe Chen, Lianrui Mu, Xiaomeng Li, Joey Tianyi Zhou, YANG FENG, Jian Wu, Haoji Hu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare Ba Con with state-of-the-art methods from two closely related fields: imbalanced semi-supervised learning and generalized category discovery. The effectiveness of Ba Con is demonstrated with superior performance over all baselines and comprehensive analysis across various datasets.
Researcher Affiliation Collaboration 1Zhejiang University, 2The Hong Kong University of Science and Technology, 3Centre for Frontier AI Research, A*STAR, 4Angelalign Technology Inc.
Pseudocode Yes The pseudo-code of Ba Con can be found in Appendix B. Algorithm 1 The overall pipeline of Ba Con. Algorithm 2 The test stage strategy of Ba Con.
Open Source Code Yes Our code is available at: https://github.com/Jianhong Bai/Ba Con.
Open Datasets Yes We conduct experiments on four popular datasets. CIFAR-10-LT/CIFAR-100-LT are long-tail subsets sampled from the original CIFAR10/CIFAR100 [10]. Image Net-100-LT is proposed by [27] with 12K images sampled from Image Net-100 [57]... Places-LT [40] contains about 62.5K images sampled from the large-scale scene-centric Places dataset [80]...
Dataset Splits No The paper describes training and testing sets, but does not explicitly mention a separate validation dataset split with proportions or counts for hyperparameter tuning or model selection.
Hardware Specification Yes We implement all our techniques using Py Torch [47] and conduct the experiments using a RTX3090 GPU.
Software Dependencies No The paper mentions 'Py Torch [47]' but does not specify its version or any other software dependencies with their respective version numbers.
Experiment Setup Yes We train with a batch size of 256, and an initial learning rate of 0.1 decayed with a cosine schedule. We train for 200 epochs on each dataset. The temperature τ is set to 1.0, the hyper-parameter p and k is set to 0.5, the sampling rates α = 0.8, β = 0.5, and the re-estimate interval r is 10 (epochs).