Beyond Class-Conditional Assumption: A Primary Attempt to Combat Instance-Dependent Label Noise

Authors: Pengfei Chen, Junjie Ye, Guangyong Chen, Jingwei Zhao, Pheng-Ann Heng11442-11450

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on MNIST and CIFAR-10 under IDN with varying noise fractions generated by Algorithm 1. The superior performance of SEAL is verified on extensive experiments, including synthetic/real-world datasets under IDN of different noise fractions, and the large benchmark Clothing1M.
Researcher Affiliation Collaboration 1 The Chinese University of Hong Kong 2 VIVO AI Lab 3 Shenzhen Key Laboratory of Virtual Reality and Human Interaction Technology, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Pseudocode Yes Algorithm 1 IDN Generation. Algorithm 2 An iteration of SEAL.
Open Source Code Yes Our code is released1. 1https://github.com/chenpf1025/IDN
Open Datasets Yes We conduct experiments on MNIST and CIFAR-10 (Krizhevsky and Hinton 2009) with varying IDN fractions as well as large-scale real-world noise benchmark Clothing1M (Xiao et al. 2015).
Dataset Splits Yes The training set consists of 1M noisy instances and the additional validation, testing sets consist of 14K, 10K clean instances. We keep 500k random samples from validation while train a Res Net-50 on the rest samples.
Hardware Specification No The paper mentions training models but does not provide specific details on the hardware used, such as GPU/CPU models or memory specifications.
Software Dependencies No The paper specifies optimizers and hyperparameters (e.g., 'SGD with a momentum of 0.5'), but it does not list specific software libraries or frameworks with their version numbers (e.g., PyTorch 1.9, TensorFlow 2.x).
Experiment Setup Yes On MNIST, ...Models are trained for 50 epochs with a batch size of 64 and we report the testing accuracy at the last epoch. For the optimizer, we use SGD with a momentum of 0.5, a learning rete of 0.01, without weight decay. On CIFAR-10, ...Models are trained for 150 epochs with a batch size of 128... For the optimizer, we use SGD with a momentum of 0.9 and a weight decay of 5 10 4. The learning rate is initialized as 0.1 and is divided by 5 after 60 and 120 epochs. ...On Clothing1M, ...Models are trained for 10 epochs with a batch size of 256... For the optimizer, we use SGD with a momentum of 0.9 and a weight decay of 10 3. We use a learning rate of 10 3 in the first 5 epochs and 10 4 in the second 5 epochs...