Overview / Usage

The annotations provided by human labelers on websites such as Amazon Mechanical Turk would introduce biases and incorrect labels. As annotating large-scale clean and unbias dataset is expensive and time-consuming, many efforts have been made to improve the robustness of ConvNets trained on noisy datasets.

Methodology / Approach

Rather than simply backpropagating the loss to update weights, we somehow need to make the sense of clean data and noisy data present in dataset, and try to make our model learn information only from the clean data.

3 Types of Approaches used to learn on Noisy data:

• The transition matrix was widely used to capture the transition probability between the noisy label and true label, i.e. the sample with a true label y has a certain probability to be mislabeled as a noisy label y˜.
  These methods worked well on the noisy dataset that is created intentionally by human with label flipping such as the noisy version of CIFAR10. However, when applying these approaches to real-world datasets it failed.
• Robust loss function against label noises. Particularly truncated loss or the loss which benefits from both mean absolute error and cross-entropy loss was proposed for training data with label noise in it. However, these robust loss functions didn’t work for every case, and weren’t so robust afterall.
• Use of an additional network to decide whether a label is noisy or not. Methods based on curriculum learning, meta-learning, co-teaching etc, were proposed. However, most of these approaches either required extra clean samples as additional information or adopted a complicated training procedure.

We propose to design a network which is self-sufficient and robust to learn the clean data distribution for classification despite of the label noise present. One of the possible approaches rely on unsupervised learning and making sense of the data via clustering.

Other links

• List of related papers (https://github.com/subeeshvasu/Awesome-Learning-with-Label-Noise)