CS-Isolate: Extracting Hard Confident Examples by Content and Style Isolation
Authors: Yexiong Lin, Yu Yao, Xiaolong Shi, Mingming Gong, Xu Shen, Dong Xu, Tongliang Liu
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we introduce the setting of our experiments and compare our experimental results with existing methods. Most of our experiments are left out in Appendix C due to the limited space. |
| Researcher Affiliation | Collaboration | 1The University of Sydney; 2Mohamed bin Zayed University of Artificial Intelligence; 3Carnegie Mellon University; 4The University of Melbourne; 5Alibaba DAMO Academy; 6The University of Hong Kong. |
| Pseudocode | Yes | Algorithm 1 CS-Isolate-DM |
| Open Source Code | Yes | The implementation is available at https://github.com/tmllab/ 2023_Neur IPS_CS-isolate. |
| Open Datasets | Yes | We evaluate our methods on three synthetic noise datasets Fashion MNIST [54], CIFAR-10 [25], and CIFAR-100 [25], and two real-world label-noise datasets, CIFAR-10N [48] and Clothing1M [55]. |
| Dataset Splits | Yes | Fashion MNIST includes 70,000 images of size 24 24, categorized into 10 classes with 60,000 for training and 10,000 for testing. Both CIFAR-10 and CIFAR-100 contain 50,000 training images and 10,000 test images; |
| Hardware Specification | Yes | We implemented our method using Py Torch 1.12.1 and performed experiments on the NVIDIA Tesla V100. |
| Software Dependencies | Yes | We implemented our method using Py Torch 1.12.1 and performed experiments on the NVIDIA Tesla V100. |
| Experiment Setup | Yes | For experiments on synthesis datasets, the initial learning rate for SGD was set at 0.02 and for Adam at 0.001. The batch size is 64. For experiments on Fashion-MNIST, our network was trained for 100 epochs. Both learning rates were reduced by a factor of 10 after 80 epochs. For experiments on CIFAR-10, CIFAR-100, and CIFAR-10N, our network was trained for 300 epochs. Both learning rates were reduced by a factor of 10 after 150 epochs. For experiments on Clothing1M, our network was trained for 80 epochs with a batch size of 32. The initial learning rate for SGD was set at 0.002 and for Adam at 0.001. Both learning rates were reduced by a factor of 10 after 40 epochs. The encoder qϕc and the classifier head fψ were warmed up on noisy data for 10 epochs for CIFAR-10 and CIFAR-10N, warmed up for 30 epochs for CIFAR-100, 5 epochs for Fashion-MNIST, and 1 epoch for Clothing1M. The dimensions of Zc and Zs were set at 32. The data augmentation techniques include shift scale rotation, random crop and horizontal flip, random brightness contrast, color jitter, and random to gray. For all experiments, we set M = 1000, λr = 1 and λELBO = 1e 3. λref was increased gradually then kept constant at 1e 3 after 140 epochs. For synthetic datasets, we set λu as 0 and 15 for the noise rates of 0.2 and 0.4 in Fashion MNIST and CIFAR-10 datasets and as 100 in CIFAR-100. For CIFAR-10N, we set λu as 50 in the 'worst' noise type and 0 in the rest of the noise types. For Clothing1M, we set λu as 0. |