Instance-dependent Label-noise Learning under a Structural Causal Model
Authors: Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, Kun Zhang
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, the proposed method outperforms all state-of-the-art methods on both synthetic and real-world label-noise datasets. In Section 4, we compare the classification accuracy of proposed method with popular label-noise learning algorithms [15, 19, 11, 10, 28, 36, 16] on both synthetic and real-world datasets. |
| Researcher Affiliation | Collaboration | Yu Yao1 Tongliang Liu1 Mingming Gong2 Bo Han3 Gang Niu4 Kun Zhang5 1TML Lab, University of Sydney; 2University of Melbourne; 3Hong Kong Baptist University; 4RIKEN AIP; 5Carnegie Mellon University |
| Pseudocode | Yes | Algorithm 1 Causal NL |
| Open Source Code | No | The paper does not provide an explicit statement or link to its open-source code. |
| Open Datasets | Yes | We examine the efficacy of our approach on manually corrupted versions of four datasets, i.e., Fashion MNIST [31], SVHN [18], CIFAR10, CIFAR100 [13], and one real-world noisy dataset, i.e., Clothing1M [32]. |
| Dataset Splits | No | The paper mentions '50,000 training images and 10,000 test images' for CIFAR10/100, but does not explicitly state training/validation/test splits, or reference predefined splits for reproducibility. |
| Hardware Specification | Yes | For a fair comparison, all experiments are conducted on NVIDIA Tesla V100, and all methods are implemented by Py Torch. |
| Software Dependencies | No | The paper mentions 'all methods are implemented by Py Torch' but does not provide specific version numbers for PyTorch or other software dependencies. |
| Experiment Setup | Yes | Dimension of the latent representation Z is set to 25 for all synthetic noisy datasets. For encoder networks ˆqφ1 1(X) and ˆqφ2 1(X), we use the same network structures with the baseline method. Specially, we use a Res Net-18 network for Fashion MNIST, a Res Net-34 network for SVHN and CIFAR10, a Res Net-50 network for CIFAR100 without pretraining. For Clothing1M, we use Res Net-50 networks pre-trained on Image Net. The data-augmentation methods random crop and horizontal flip are used for our method. For Clothing1M, we use a Res Net-50 network pre-trained on Image Net, and the clean training data is not used. The dimensionality of the latent representation Z is set to 100. |