Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical

Authors: Wei Wang, Takashi Ishida, Yu-Jie Zhang, Gang Niu, Masashi Sugiyama

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on both synthetic and real-world benchmark datasets validate the superiority of our proposed approach over state-of-the-art methods.
Researcher Affiliation Academia 1The University of Tokyo 2RIKEN.
Pseudocode Yes Algorithm 1 SCARCE
Open Source Code Yes Our implementation of SCARCE is available at https://github.com/wwangwitsel/SCARCE.
Open Datasets Yes We conducted experiments on synthetic benchmark datasets, including MNIST (Le Cun et al., 1998), Kuzushiji-MNIST (Clanuwat et al., 2018), Fashion-MNIST (Xiao et al., 2017), and CIFAR-10 (Krizhevsky & Hinton, 2009).
Dataset Splits Yes The training curves and test curves of the method that works by minimizing the URE in Eq. (9) are shown in Figure 1.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models used for running the experiments.
Software Dependencies No All the methods were implemented in Py Torch (Paszke et al., 2019). We used the Adam optimizer (Kingma & Ba, 2015).
Experiment Setup Yes The learning rate and batch size were fixed to 1e-3 and 256 for all the datasets, respectively. The weight decay was 1e-3 for CIFAR-10 and 1e-5 for the other three datasets. The number of epochs was set to 200, and we recorded the mean accuracy in the last ten epochs.