Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning

Authors: Dong-Dong Wu, Deng-Bao Wang, Min-Ling Zhang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments and analysis on multiple datasets validate the rationality and superiority of our proposed approach.
Researcher Affiliation Academia School of Computer Science and Engineering, Southeast University, Nanjing 210096, China Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, China {dongdongwu, wangdb, zhangml}@seu.edu.cn
Pseudocode Yes The pseudo-code of our complete algorithm DIRK is shown in Appendix A.3. ... Thoroughly, the pseudo-code and flowchart of DIRK-REF is shown in Appendix A.4.
Open Source Code Yes Source code is available at https://github.com/wu-dd/DIRK.
Open Datasets Yes We evaluated our method on seven commonly used benchmark image dataset: Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017), Kuzushiji-MNIST (Clanuwat et al. 2018), CIFAR-10 (Krizhevsky, Hinton et al. 2009), CIFAR100 (Krizhevsky, Hinton et al. 2009), CUB-200 (Welinder et al. 2010), Flower (Nilsback and Zisserman 2008) and Oxford-IIIT Pet (Parkhi et al. 2012).
Dataset Splits Yes The hyperparameters were selected so as to maximize the accuracy on a validation set (10% of the training set).
Hardware Specification Yes Our implementation was executed using Py Torch (Paszke et al. 2019), and all experiments were conducted with NVIDIA Tesla V100 GPU.
Software Dependencies No Our implementation was executed using Py Torch (Paszke et al. 2019). While PyTorch is mentioned, a specific version number is not provided.
Experiment Setup Yes For all methods on benchmark datasets, we used SGD as the optimizer with a momentum of 0.9, a weight decay of 1e3, an initial learning rate of 1e-2, and set the epoch number to 500. ... We set the momentum hyperparameter m as 0.99 and the trade-off parameter λ as 0 in DIRK. ... temperature hyperparameters τ1 = 0.01, τ2 = 0.07, and the sizes of both queues are fixed to be 1024. For large-scale datasets CUB-200, Flower, and Oxford-IIIT Pet, we set the mini-batch size as 32, while 256 for other datasets.