Figure - available via license: Creative Commons Attribution 4.0 International
Content may be subject to copyright.
Accuracy (↑) on two clean and noisy datasets (Wei et al., 2022). The two best results are highlighted in bold. Note that NR stands for Noise Rate. CIFAR-10N includes three annotations per image, and Noisy split denotes various aggregation strategies. Following the proposed naming convention, Aggr refers to a majority voting, R-i (i ∈ {1, 2, 3}) denotes the i-th submitted label for each image, and Worst denotes the selection of only wrong labels, if applicable. Conversely, CIFAR-100N presents just one annotation per image, thereby offering one single noisy split.
Source publication
The performance of deep neural networks scales with dataset size and label quality, rendering the efficient mitigation of low-quality data annotations crucial for building robust and cost-effective systems. Existing strategies to address label noise exhibit severe limitations due to computational complexity and application dependency. In this work,...
Contexts in source publication
Context 1
... in scenarios where text embeddings are not explicitly needed, as in our case, opting for an image-centric training paradigm seems beneficial. Results Table 2 shows the results of the designed experiment for real-world noisy labels. The introduced weighting scheme in WANN substantially enhances robustness compared to ANN, particularly in high-noise conditions. ...Context 2
... Pattern Clean Symmetric Symmetric Symmetric Clean Symmetric Asymmetric Instance NR - 20% 30% 40% - 60% 30% Table 4: Accuracy (↑) on BreastMNIST and DermaMNIST. We highlight the significantly best method (paired t-test, p < 0.05) or the top two methods if their difference is not statistically significant. ...