Timezone: »

Guaranteed Robust Deep Learning against Extreme Label Noise using Self-supervised Learning
Yihao Xue · Kyle Whitecross · Baharan Mirzasoleiman

@ None #None

Self-supervised contrastive learning has been recently shown very effective in preventing deep networks from overfitting noisy labels. Despite its empirical success, the theoretical understanding of the effect of contrastive learning on boosting robustness of deep networks is very limited. In this work, we show that contrastive learning provably boosts robustness of deep networks against noisy labels by providing an embedding matrix that has (i) a singular value corresponding to each subclass in the data, which is relatively larger than the sum of the remaining singular values of that subclass; and (ii) a large alignment between the largest singular vector and the clean labels of that subclass. The above properties allow a linear layer trained on the embeddings to learn the clean labels quickly, and prevent it from overfitting the noisy labels for a large number of training iterations.We further show that the initial robustness provided by contrastive learning enables state-of-the-art robust methods to achieve a superior performance under extreme noise levels, e.g., 6.3\% increase in accuracy on CIFAR-10 with 40\% asymmetric noisy labels, and 14\% increase in accuracy on CIFAR100 with 80\% symmetric noisy labels.

Author Information

Yihao Xue (UCLA)
Kyle Whitecross (UCLA)

UCLA undergrad studying CS and ML intern at Kumo.ai.

Baharan Mirzasoleiman (Stanford University)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors