Delving into Sample Loss Curve to Embrace Noisy and Imbalanced Data
Authors: Shenwang Jiang, Jianan Li, Ying Wang, Bo Huang, Zhang Zhang, Tingfa Xu7024-7032
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive synthetic and real experiments well validate the proposed method, which achieves state-of-the-art performance on multiple challenging benchmarks. We use CIFAR dataset (Krizhevsky, Hinton et al. 2009) with varying noise rates and imbalance ratios to verify the effectiveness of the propose method. We also test our method on real noisy and imbalanced data to validate its generality. |
| Researcher Affiliation | Academia | 1 Beijing Institute of Technology 2 University of Massachusetts Lowell {jiangwenj02, lijianan15, wangying7275}@gmail.com, a1039377853@163.com Zhang Zhang@student.uml.edu, ciom xtf1@bit.edu.cn |
| Pseudocode | No | The paper describes the methods and equations but does not provide a structured pseudocode or algorithm block. |
| Open Source Code | Yes | Code is available at https://github.com/jiangwenj02/Curve Net-V1 |
| Open Datasets | Yes | We use CIFAR dataset (Krizhevsky, Hinton et al. 2009) with varying noise rates and imbalance ratios to verify the effectiveness of the propose method. CIFAR-10. This dataset consists of 60,000 RGB images (50,000 for training and 10,000 for testing). CIFAR-100. This dataset comprises 100 categories, each of which contains 600 images. Clothing1M. This dataset (Xiao et al. 2015) comprises 1M clothing images of 14 categories crawled from online shopping websites. Food-101N. This dataset (Lee et al. 2018) is a large-scale dataset (310k/25k training/test images) accompanied by 55k images with clean verification labels. |
| Dataset Splits | Yes | We randomly select 100 images from each category to form unbiased meta data set. CIFAR-100. This dataset comprises 100 categories, each of which contains 600 images. We randomly select 10 images from each category to form our unbiased meta-data set. |
| Hardware Specification | Yes | We train the model for 200 epochs on a single NVIDIA GTX 1080Ti. |
| Software Dependencies | No | The paper mentions using SGD and Adam optimizers, but does not provide specific version numbers for any software dependencies like programming languages or libraries (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | During allocating stage, we use stochastic gradient descent (SGD) with initial learning rate 0.1 and decrease the learning rate to 0.01 and 0.001 at epoch 80 and 100, respectively. We use a batchsize of 128 images. Curve Net is optimized using Adam with learning rate 0.001. ...cyclical learning rate (Smith 2017) is adopted to train the classifier F(ω) in the probing stage. |