Abstract:Deep Learning systems have shown tremendous accuracy in image classification, at the cost of big image datasets. Collecting such amounts of data can lead to labelling errors in the training set. Indexing multimedia content for retrieval, classification or recommendation can involve tagging or classification based on multiple criteria. In our case, we train face recognition systems for actors identification with a closed set of identities while being exposed to a significant number of perturbators (actors unknown to our database). Face classifiers are known to be sensitive to label noise. We review recent works on how to manage noisy annotations when training deep learning classifiers, independently from our interest in face recognition.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how deep - learning classifiers deal with data sets with noisy labels during the training process. Specifically, the author focuses on how to effectively train deep - learning models when there are a large number of mis - labeled samples in the data set to ensure that their performance and accuracy are not affected. ### Problem Background When constructing large - scale image data sets, due to the wide and diverse data sources, mis - labeling is inevitable. These noisy labels will have a negative impact on the model training, leading to a decline in model performance. Especially in tasks such as face recognition, the impact of noisy labels is particularly significant. Therefore, how to deal with these noisy labels has become an important research topic. ### Main Problems 1. **The impact of noisy labels on model performance**: Noisy labels can mislead the model's learning process, causing the model to over - fit the noise or fail to generalize correctly. 2. **How to detect and correct noisy labels**: Effective techniques need to be developed to identify and correct noisy labels in the data set. 3. **How to design robust learning algorithms**: It is still possible to train high - performance classifiers even in the presence of noisy labels. ### Solution Overview The paper reviews the research progress on the noisy label problem in recent years and proposes several main solutions: 1. **Prediction Reweighting**: Adjust the weights of model predictions by estimating the confusion matrix, thereby reflecting the uncertainty of each observed label. 2. **Sample Importance Reweighting**: Re - adjust the importance of samples in training according to the probability of samples being contaminated by noise. 3. **Unlabeling**: Treat samples considered to be noisy as unlabeled samples and use semi - supervised or unsupervised methods for training. 4. **Label Fixing**: Try to directly correct noisy labels so that the model can be trained on the correct labels. ### Experimental Verification To verify the effectiveness of these methods, the author uses multiple public data sets (such as CIFAR - 10, MNIST, Clothing1M, Food101 - N, and WebVision) and conducts experiments by introducing artificial noisy labels. The experimental results show that different methods perform differently under different types of noise, each having its own advantages and disadvantages. ### Conclusion This paper not only summarizes the existing noisy label processing methods but also points out the application scenarios and limitations of each method. Future research can further explore how to combine multiple methods to better deal with the complex and diverse noisy label problems.

Deep Learning Classification With Noisy Labels

FGCM: Noisy Label Learning via Fine-Grained Confidence Modeling

Noisy Label Processing for Classification: A Survey

Robust Image Classification with Noisy Labels by Negative Learning and Feature Space Renormalization

Noisy Label Learning for Large-scale Medical Image Classification

Image Classification with Deep Learning in the Presence of Noisy Labels: A Survey

LABELNET: Recovering Noisy Labels

Learning with Noisy Labels Via Self-supervised Adversarial Noisy Masking

Deep learning with noisy labels: Exploring techniques and remedies in medical image analysis

Model-agnostic Approaches to Handling Noisy Labels When Training Sound Event Classifiers

Deep Image Retrieval is not Robust to Label Noise

Learning Image Labels On-the-fly for Training Robust Classification Models

Decoding class dynamics in learning with noisy labels

Learning Sound Event Classifiers from Web Audio with Noisy Labels

Deep Learning From Noisy Image Labels With Quality Embedding

Training Deep Neural Networks for Image Applications with Noisy Labels by Complementary Learning

Uncertainty Based Detection and Relabeling of Noisy Image Labels

A Fundus Image Classification Framework for Learning with Noisy Labels.

Learning from Noisy Labels with Noise Modeling Network

NoisyAG-News: A Benchmark for Addressing Instance-Dependent Noise in Text Classification

Towards Understanding Deep Learning from Noisy Labels with Small-Loss Criterion