Exploiting Class Activation Value for Partial-Label Learning

Authors: Fei Zhang, Lei Feng, Bo Han, Tongliang Liu, Gang Niu, Tao Qin, Masashi Sugiyama

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on various datasets demonstrate that our proposed CAVL method achieves stateof-the-art performance. In this section, extensive experiments on various datasets are implemented to verify the effectiveness and rightness of our proposed CAV and CAVL method.
Researcher Affiliation Collaboration 1Shanghai Jiao Tong University 2Hong Kong Baptist University 3Chongqing University 4The University of Sydney 5RIKEN 6Microsoft Research Asia 7The University of Tokyo
Pseudocode Yes Algorithm 1 CAVL Algorithm
Open Source Code No The paper states 'This work is partially supported by Huawei Mind Spore (Huawei, 2020).' and provides a URL 'https://www.mindspore.cn/'. This refers to a general platform, not the specific source code for the methodology described in the paper. No other explicit statement or link providing access to their source code was found.
Open Datasets Yes We use four popular benchmark datasets to test the performance of our CAVL, which are MNIST (Le Cun et al., 1998), Fashion-MNIST (Xiao et al., 2017), Kuzushiji MNIST (Clanuwat et al., 2018) and CIFAR-10 (Krizhevsky et al., 2009).
Dataset Splits Yes Hyper-parameters are searched to maximize the accuracy on a validation set containing 10% of the training samples annotated by true labels.
Hardware Specification Yes All the implemented methods are trained on 1 RTX3090 GPU with 24 GB memory.
Software Dependencies No The paper mentions 'Huawei Mind Spore (Huawei, 2020)' as a supporting framework, but does not provide specific version numbers for it or any other software libraries or dependencies used in the experiments.
Experiment Setup Yes For all methods we search the initial learning rate from {0.0001, 0.001, 0.01, 0.05, 0.1, 0.5}, and weight decay from {10-6, 10-5, ..., 10-1}. We take a mini-batch size of {32, 256} images and train all the methods using Adam (Kingma & Ba, 2015) optimizer for 250 epochs.