Learning with Biased and Noisy Labels

Recent advances in Artificial Intelligence (AI) have been built on large scale datasets. These advances often come with increasing demands on labeling, which are expensive and time consuming. Therefore, AI tends to develop its higher-level intelligence like human to capture knowledge from cheap but weak supervision such as the mislabeled data. However, current AI suffers from severely degraded performance on noisily labeled data. Thus, it is a compelling demand to design novel algorithms to enable AI to learn from noisy labels. Label noise methods such as robust loss functions assume that a fraction of data is correctly labeled to ensure effective learning. When all labels are incorrect, they often fail due to severe bias and noises. Here, we consider a kind of incorrect label, complementary label which specify a class that a feature do not belong to. We propose a general method to modify loss functions such that the classifier learned from biased complementary labels can be identical to the optimal one learned from true labels. Another challenge in label noise is the shift between distributions of training (source) and test (target) data. Existing methods often ignore these changes and they cannot learn transferable knowledge across domains. Therefore, we propose a novel Denoising Conditional Invariant Component framework which provably ensures identification of invariant representations and label distribution of target data given examples with noisy labels in source domain and unlabeled examples in target domain. Finally, we study how to estimate the noise rates in label noise. Previous methods deliver promising results but rely on strong assumptions. We can see, noise rate estimation is essentially a mixture proportion estimation problem. We also prove that noise rates can be uniquely identified and efficiently obtained under a weaker linear independent assumption