Robustness of Accuracy Metric and its Inspirations in Learning with Noisy Labels

Authors: Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, Pheng-Ann Heng11451-11461

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We verify our theoretical results and additional claims with extensive experiments. We show characterizations of models trained with noisy labels, motivated by our theoretical results, and verify the utility of a noisy validation set by showing the impressive performance of a framework termed noisy best teacher and student (NTS). Our code is released.
Researcher Affiliation Collaboration Pengfei Chen,1 Junjie Ye,2 Guangyong Chen,3 Jingwei Zhao,2 Pheng-Ann Heng1,3 1 The Chinese University of Hong Kong 2 VIVO AI Lab 3 Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Pseudocode No The paper describes the NTS framework steps in text but does not provide a formally labeled pseudocode or algorithm block.
Open Source Code Yes Our code is released1. 1https://github.com/chenpf1025/Robustness Accuracy
Open Datasets Yes CIFAR-10 and CIFAR-100. We use Wide Res Net-28-10 (WRN-28-10) (Zagoruyko and Komodakis 2016) as the classifier on CIFAR-10 and CIFAR-100. Clothing1M. Clothing1M is a large-scale benchmark containing real-world noise.
Dataset Splits Yes We corrupt the training set which has 50000 samples and randomly split 5000 noisy samples for validation. use 1 million noisy samples of Clothing1M for training, 14k and 10k clean data respectively for validation and test. We randomly sample 14k noisy samples (i.e., 1k samples per class) from the 1 million noisy training samples for validation.
Hardware Specification Yes All models are trained with Tesla V100 GPU.
Software Dependencies No The paper mentions software components like 'SGD optimizer', 'Cross Entropy (CE) loss', 'Generalized Cross Entropy (GCE)', 'Co-teaching (Co-T)', and 'Determinant based Mutual Information (DMI)' but does not provide specific version numbers for any of these or the underlying programming languages/frameworks.
Experiment Setup Yes CE, GCE and Co-T shares the same batch size of 128 and learning rate schedule, i.e., training using SGD optimizer for 200 epochs, with a initial learning rate of 0.1, which is decreased by a factor of 5 after 60, 120 and 160 epochs. Following its original paper and official implementation, DMI uses a model pretrained by CE as initialization and requires a larger batch size of 256 and a smaller learning rate, which is tuned in {10 4, 10 5, 10 6} and fixed to 10 6 finally. It is trained using SGD optimizer for 100 epochs without learning rate change. In all methods, the SGD optimizer is implemented with momentum 0.9 and weight decay 5 10 4.