$\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise
Authors: Jialiang Wang, Xiong Zhou, Deming Zhai, Junjun Jiang, Xiangyang Ji, Xianming Liu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the superiority of our method in mitigating synthetic and real-world label noise. |
| Researcher Affiliation | Academia | Jialiang Wang1 Xiong Zhou1 Deming Zhai1 Junjun Jiang1 Xiangyang Ji2 Xianming Liu1 1Faculty of Computing, Harbin Institute of Technology 2Department of Automation, Tsinghua University |
| Pseudocode | Yes | Algorithm 1 CEϵ+MAE (Semi) |
| Open Source Code | Yes | The code is available at https://github.com/cswjl/eps-softmax. |
| Open Datasets | Yes | We evaluate our proposed methods on benchmark datasets CIFAR-10 / CIFAR-100 [24] with synthetic label noise... We further conduct comparison studies on human-annotated datasets CIFAR-10N/CIFAR-100N [28]... We perform experiments on massively real-world noisy datasets, including Web Vision [44], ILSVRC12 (Image Net) [45] and Clothing1M [46] |
| Dataset Splits | No | The paper specifies training on the given datasets and evaluating on test sets, but does not explicitly provide details on a distinct validation set split for reproducibility beyond what might be implicitly defined by referring to prior work for experimental settings. |
| Hardware Specification | Yes | All experiments are implemented by Py Torch and are conducted on NVIDIA Ge Force RTX 4090. |
| Software Dependencies | No | All experiments are implemented by Py Torch and are conducted on NVIDIA Ge Force RTX 4090. |
| Experiment Setup | Yes | An 8-layer CNN is used for CIFAR-10 and a Res Net-34 for CIFAR-100. The networks are trained for 120 and 200 epochs for CIFAR-10 and CIFAR-100 with batch size 128. We use the SGD optimizer with momentum 0.9 and cosine learning rate annealing. The weight decay is set to 1 10 4 and 1 10 5 for CIFAR-10 and CIFAR-100. The initial learning rate is set to 0.01 for CIFAR-10 and 0.1 for CIFAR-100. |