Lower-Bounded Proper Losses for Weakly Supervised Classification
Authors: Shuhei M Yoshida, Takashi Takenouchi, Masashi Sugiyama
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Furthermore, we experimentally demonstrate the effectiveness of our proposed approach, as compared to improper or unbounded losses. The results highlight the importance of properness and lower-boundedness. |
| Researcher Affiliation | Collaboration | 1Biometrics Research Laboratories, NEC Corporation, Kawasaki, Kanagawa, Japan 2RIKEN Center for Advanced Intelligence Project, Chuo-ku, Tokyo, Japan 3National Graduate Institute for Policy Studies, Minato-ku, Tokyo, Japan 4Department of Complexity Science and Engineering, The University of Tokyo, Kashiwa, Chiba, Japan. |
| Pseudocode | Yes | Algorithm 1 Training of the linear model with the backward-corrected cross entropy and generalized logit squeezing. |
| Open Source Code | Yes | The code is publicly available at https://github.com/yoshum/lower-bounded-proper-losses. |
| Open Datasets | Yes | We evaluated the effectiveness of the losses on the MNIST (Le Cun et al., 1998) and CIFAR-10 (Krizhevsky, 2009) datasets. |
| Dataset Splits | No | The momentum was fixed to 0.9, while the initial learning rates were chosen as those giving the best validation accuracy. More details on the experimental procedure and the hyperparameters are given in Appendix G. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions PyTorch in the references but does not specify the version of PyTorch or any other software dependencies used for the experiments. |
| Experiment Setup | Yes | We used stochastic gradient descent with momentum to optimize the models. The momentum was fixed to 0.9, while the initial learning rates were chosen as those giving the best validation accuracy. The default value of the weight decay was 10 4, but we also tuned it with BC and BC + GA to compare its effect with that of g LS. |