CleanNet: Transfer Learning for Scalable Image Classifier Training with Label Noise (1711.07131v2)

Abstract: In this paper, we study the problem of learning image classification models with label noise. Existing approaches depending on human supervision are generally not scalable as manually identifying correct or incorrect labels is time-consuming, whereas approaches not relying on human supervision are scalable but less effective. To reduce the amount of human supervision for label noise cleaning, we introduce CleanNet, a joint neural embedding network, which only requires a fraction of the classes being manually verified to provide the knowledge of label noise that can be transferred to other classes. We further integrate CleanNet and conventional convolutional neural network classifier into one framework for image classification learning. We demonstrate the effectiveness of the proposed algorithm on both of the label noise detection task and the image classification on noisy data task on several large-scale datasets. Experimental results show that CleanNet can reduce label noise detection error rate on held-out classes where no human supervision available by 41.5% compared to current weakly supervised methods. It also achieves 47% of the performance gain of verifying all images with only 3.2% images verified on an image classification task. Source code and dataset will be available at kuanghuei.github.io/CleanNetProject.

• The paper shows that using transfer learning and a novel attention mechanism can reduce label noise error by 41.5% in unverified classes.
• It introduces a dual-encoder architecture that leverages reference set and query embeddings to compare class prototypes and detect mislabeled images.
• The proposed method achieves 47% of the performance gain of full manual verification while only inspecting 3.2% of images, offering scalable solutions.

CleanNet: Transfer Learning for Scalable Image Classifier Training with Label Noise

The paper focuses on the persistent challenge in machine learning of training image classifiers in the presence of noisy labels, a common occurrence when using large-scale datasets sourced from the Internet. Traditional methods of dealing with label noise often depend either on comprehensive human oversight or completely eschew any form of supervision, leading to either scalability or effectiveness issues, but rarely both. CleanNet proposes a solution that strikes a balance by utilizing a limited amount of manual verification to inform a broader, scalable approach to denoising data.

CleanNet leverages a transfer learning approach within its neural architecture, allowing the knowledge gained from a small subset of manually verified classes to be extended to unverified classes. This transfer learning capability is grounded in an innovative attention mechanism that selects "class prototypes" to represent each class effectively. CleanNet consists of a reference set encoder, which generates class embeddings, and a query encoder, which produces query embeddings. These embeddings are compared to detect potential mislabeling through a similarity constraint, offering a robust framework for label noise detection and image classification.

Performance evaluations of CleanNet reveal significant improvements in both label noise detection and image classification tasks. For instance, CleanNet achieved a 41.5% reduction in label noise detection error rate on classes without manual verification compared to existing weakly supervised methods. Additionally, it attained 47% of the performance gain of completely verifying all class labels while only verifying 3.2% of the images.

The architecture of CleanNet and the described attention mechanism suggest several practical implications. CleanNet offers a pragmatic route towards handling label noise in exercises where manual verification of data is untenable at full scale, with the framework being readily applicable to various domains beyond image classification where similar data quality challenges exist.

Theoretically, the paper enriches the dialogue surrounding transfer learning by expanding its utility to mislabeled data scenarios, broadening the potential applications of joint neural embedding networks. Future research could explore integrating CleanNet into other machine learning paradigms, potentially with adjustments to the attention mechanisms or embedding constraints to better accommodate varying types of data and classification challenges.

In closing, CleanNet represents a significant step towards scalable and effective machine learning methods in the presence of label noise, suggesting new paths for research and application in artificial intelligence, particularly in domains requiring extensive data cleansing mechanisms without exhaustive human oversight.