Learning from Bad Data via Generation

Authors: Tianyu Guo, Chang Xu, Boxin Shi, Chao Xu, Dacheng Tao

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate our methods on three kinds of bad data environments: (i) long-tailed training set classification on the MNIST [25], FMNIST [42], and CIFAR-10 [23] datasets; (ii) classification of distorted test set on the CIFAR-10 and SVHN [33] datasets; and (iii) reduced training set generation task on the FMNIST and CIFAR-10 datasets. We resize images in the MNIST and FMNIST datasets to 32 32 for convenience. Moreover, we use a conditional version of WGANGP [15] on all datasets except the CIFAR-10 datasets on which we use the 32 resolution version of Big GAN [4] instead. The classifier implemented on the MNIST and FMNIST has comparable architecture to Triple GAN [27], and we use VGG-16 [38] and Res Net-101 [18] on the CIFAR-10 and SVHN datasets. We implement our experiments based on Py Torch. For generator and discriminator we use a learning rate of 0.0002, while 0.02 is for the classifier, the learning rate decay is deployed, and the optimizer is Adam. Experiments are conducted on 4 NVIDIA 1080Ti GPUs.
Researcher Affiliation Collaboration Tianyu Guo1,2, , Chang Xu2, , Boxin Shi3,4, , Chao Xu1, Dacheng Tao2 1Key Laboratory of Machine Perception (MOE), CMIC, School of EECS, Peking University, 100871, China 2UBTECH Sydney AI Centre, School of Computer Science, Faculty of Engineering, The University of Sydney, Darlington, NSW 2008, Australia 3National Engineering Laboratory for Video Technology, Department of Computer Science and Technology, Peking University, Beijing, 100871, China 4Peng Cheng Laboratory, Shenzhen, 518040, China {tianyuguo, shiboxin}@pku.edu.cn, chaoxu@cis.pku.edu.cn {c.xu, dacheng.tao}@sydney.edu.au
Pseudocode Yes Our proposed algorithm is summarized in Algorithm 1 in Appendix.
Open Source Code No No statement explicitly providing access to the source code for the described methodology was found.
Open Datasets Yes In this section, we evaluate our methods on three kinds of bad data environments: (i) long-tailed training set classification on the MNIST [25], FMNIST [42], and CIFAR-10 [23] datasets; (ii) classification of distorted test set on the CIFAR-10 and SVHN [33] datasets;
Dataset Splits No The paper does not explicitly state the training/validation/test dataset splits in terms of percentages, counts, or by referencing standard splits with specific details. It mentions using 'training set' and 'test set' but no explicit splits.
Hardware Specification Yes Experiments are conducted on 4 NVIDIA 1080Ti GPUs.
Software Dependencies No The paper mentions 'Py Torch' but does not provide specific version numbers for it or any other software libraries or dependencies.
Experiment Setup Yes For generator and discriminator we use a learning rate of 0.0002, while 0.02 is for the classifier, the learning rate decay is deployed, and the optimizer is Adam.