Training Deep Neural Networks on Noisy Labels with Bootstrapping

TL;DR

This paper introduces a novel bootstrapping method that enhances deep neural network training by making predictions consistent with similar percepts, significantly improving robustness to noisy and incomplete labels across various visual recognition tasks.

Contribution

The authors propose a generic consistency-based augmentation to the training objective, improving deep learning robustness to label noise and subjectivity without requiring additional modifications.

Findings

Robustness to label corruption demonstrated on MNIST.

Achieved state-of-the-art results on emotion recognition with subjective labels.

Improved detection performance on ILSVRC2014 with deep networks.

Abstract

Current state-of-the-art deep learning systems for visual object recognition and detection use purely supervised training with regularization such as dropout to avoid overfitting. The performance depends critically on the amount of labeled examples, and in current practice the labels are assumed to be unambiguous and accurate. However, this assumption often does not hold; e.g. in recognition, class labels may be missing; in detection, objects in the image may not be localized; and in general, the labeling may be subjective. In this work we propose a generic way to handle noisy and incomplete labeling by augmenting the prediction objective with a notion of consistency. We consider a prediction consistent if the same prediction is made given similar percepts, where the notion of similarity is between deep network features computed from the input data. In experiments we demonstrate that our approach yields substantial robustness to label noise on several datasets. On MNIST handwritten digits, we show that our model is robust to label corruption. On the Toronto Face Database, we show that our model handles well the case of subjective labels in emotion recognition, achieving state-of-the- art results, and can also benefit from unlabeled face images with no modification to our method. On the ILSVRC2014 detection challenge data, we show that our approach extends to very deep networks, high resolution images and structured outputs, and results in improved scalable detection.