Estimating Instance-dependent Bayes-label Transition Matrix using a Deep Neural Network
Authors: Shuo Yang, Erkun Yang, Bo Han, Yang Liu, Min Xu, Gang Niu, Tongliang Liu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of the proposed method is verified on three synthetic noisy datasets and a large-scale realworld noisy dataset, significant performance improvements on both synthetic and real-world noisy datasets and all experiment settings are achieved. |
| Researcher Affiliation | Academia | 1University of Technology Sydney 2Xidian University 3Hong Kong Baptist University 4Computer Science and Engineering, UC Santa Cruz 5RIKEN Center for Advanced Intelligence Project 6TML Lab, Sydney AI Centre, The University of Sydney. |
| Pseudocode | Yes | Algorithm 1 Instance-dependent Label-noise Learning with Bayes Label Transition Network. |
| Open Source Code | No | The paper does not explicitly provide a link to its source code or state that it will be made publicly available. |
| Open Datasets | Yes | We conduct the experiment on four datasets to verify the effectiveness of our proposed method, where three of them are manually corrupted, i.e., F-MNIST, CIFAR-10, and SVHN, one of them is real-world noisy datasets, i.e., Clothing1M. |
| Dataset Splits | Yes | F-MNIST has 28 28 grayscale images of 10 classes including 60,000 training images and 10,000 test images. CIFAR-10 dataset contains 50,000 color images from 10 classes for training and 10,000 color images from 10 classes for testing both with shape of 32 32 3. SVHN has 10 classes of images with 73,257 training images and 26,032 test images. [...] 10% of the noisy training examples of all datasets are left out as a noisy validation set for model selection. |
| Hardware Specification | Yes | All the codes are implemented in Py Torch 1.6.0 with CUDA 10.0, and run on NVIDIA Tesla V100 GPUs. |
| Software Dependencies | Yes | All the codes are implemented in Py Torch 1.6.0 with CUDA 10.0, and run on NVIDIA Tesla V100 GPUs. |
| Experiment Setup | Yes | We first use SGD with momentum 0.9, batch size 128, and an initial learning rate of 0.01 to warm up the network for five epochs on the noisy dataset. The classification network is trained on the noisy dataset for 50 epochs for F-MNIST, CIFAR-10 and SVHN and for 10 epochs for Clothing1M using Adam optimizer with a learning rate of 5e 7 and weight decay of 1e 4. |