Learning from Noisy Labels with Complementary Loss Functions
Authors: Deng-Bao Wang, Yong Wen, Lujia Pan, Min-Ling Zhang10111-10119
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on benchmark classification datasets indicate that the proposed method helps achieve robust and sufficient deep neural network training simultaneously. |
| Researcher Affiliation | Collaboration | 1School of Computer Science and Engineering, Southeast University, Nanjing 210096, China 2Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, China 3Noah s Ark Lab, Huawei Technologies 4NSKEYLAB, Xi an Jiaotong University 5Collaborative Innovation Center of Wireless Communications Technology, China |
| Pseudocode | Yes | Algorithm 1: Learning with Complementary Losses. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described, nor does it explicitly state that the code is open-source or publicly available. |
| Open Datasets | Yes | To verify the superiority of our approach, we conduct experiments on two commonly used image classification datasets in the literature of noise-label learning: CIFAR-10 and CIFAR-100, consisting of 32x32 color images arranged in 10 and 100 classes, respectively. Both datasets contain 50,000 training and 10,000 test images. We further use Tiny Image Net (subset of Image Net (Deng et al. 2009)) to test the generality of our approach. Tiny Image Net contains 200 classes with 100K training images, 10K validation ones with resolution 64x64. In addition, we also conduct experiment on Clothing1M, which contains 14 classes with 1M real-world noisy training samples. |
| Dataset Splits | Yes | Both datasets [CIFAR-10/100] contain 50,000 training and 10,000 test images. We further use Tiny Image Net (subset of Image Net (Deng et al. 2009)) to test the generality of our approach. Tiny Image Net contains 200 classes with 100K training images, 10K validation ones with resolution 64x64. |
| Hardware Specification | Yes | The implementation is based on Py Torch (Paszke et al. 2019) and experiments were carried out with NVIDIA Tesla V100 GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch (Paszke et al. 2019)' as the basis for implementation but does not provide specific version numbers for PyTorch or any other software dependencies, which are required for full reproducibility. |
| Experiment Setup | Yes | We use Pre Act-Res Net-18 and train it using SGD with a momentum of 0.9, a weight decay of 0.0005, and a batch size 100 in our experiments. For both CIFAR-10 and CIFAR-100, the network is trained for 300 epochs in which the first 60 epochs are for warming up the networks. In warm-up stage, we use the weighted gradient as shown in Eq.(6) for parameter updating. We set the initial learning rate as 0.02 and reduce it by a factor of 10 after 200 epochs. For other hyperparameters of our method, we simply set α = 1, γ1 = 0.9 and γ2 = 1 for all cases. For Tiny Image Net, the total number of epochs is 200, and the initial learning rate is reduced by a factor of 10 after 100 epochs. |