An Information Fusion Approach to Learning with Instance-Dependent Label Noise

Authors: Zhimeng Jiang, Kaixiong Zhou, Zirui Liu, Li Li, Rui Chen, Soo-Hyun Choi, Xia Hu

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations on synthetic and real-world datasets demonstrate that our method is superior to the state-of-the-art approaches, and achieves more stable training for instance-dependent label noise.
Researcher Affiliation Collaboration Zhimeng Jiang1, Kaixiong Zhou2, Zirui Liu2, Li Li3, Rui Chen3, Soo-Hyun Choi4 , Xia Hu2 1Texas A&M University, 2Rice University, 3Samsung Research America, 4Samsung Electronics
Pseudocode Yes Algorithm 1: Information Fusion Algorithm Input :Noisy dataset D = (xn, yn)N n=1; Noisy validation data; tolerant epochs t. Output :The robust neural network over noisy label. Algorithm 2: Instance-Dependent Label Noise Generation Input :Clean samples (xn, yn)N n=1; Overall noise rate τ; Number of classes c; Size of input features: 1 d. Output :Noisy samples (xi, yn)N n=1.
Open Source Code No The paper does not provide a concrete access link or explicit statement about the release of its source code.
Open Datasets Yes Datasets. We verify the superiority of IF on three manually corrupted datasets, i.e., F-MNIST, SVHN, CIFAR-10, and one real-world noisy dataset Clothing1M. The first three datasets contain clean data, and we manually corrupt the labels of the training datasets by following (Xia et al., 2020). IDN-τ means that the controlled noise rate is τ. All experiments on those datasets with synthetic instance-dependent label noise are repeated five times. The real-world dataset Clothing1M has 1M images with real-world noisy labels and 10k images with clean labels for testing.
Dataset Splits Yes In the experiments, we leave out 10% of the noisy training samples as a noisy validation set for model selection.
Hardware Specification Yes All baselines and IF approaches are implemented in Py Torch, and tested on a machine with AMD EPYC 7282 16-core processors, 4 Ge Force GTX-3090 Ti GPUs with 24GB memory size.
Software Dependencies No The paper mentions that the methods are "implemented in Py Torch" but does not specify the version number of PyTorch or any other software dependencies.
Experiment Setup Yes Specifically, we use a Res Net-18 network for F-MNIST, a Res Net-34 network for SVHN and CIFAR-10. For the optimization, we first use SGD with 0.9 momentum, 10 4 weight decay, 128 batch size, 50 epochs and an initial learning rate of 10 2 to initialize the network. Then, we adopt Adam optimizer and 5 10 7 learning rate to learn the NTM following PTD (Xia et al., 2020).