This survey reviews methods to counter noisy labels in image classification, indicating the importance of accurate data labeling.
Key Points
Significant damage to model training occurs due to incorrect labels, known as noisy labels, which are unavoidable in data annotation.
Different deep learning approaches have evolved to combat noisy labels, crucial for successful image classification tasks.
Exploration of real-world label noise patterns led to a synthetic benchmark using the CIFAR-10N dataset, enhancing classification robustness testing methods involved in dataset evaluation and algorithm design aspects of neural networks, which are essential for improving model accuracy.