Combining PENCIL with AMDIM for image classification with noisy and sparsely labeled data
Alexandros Petropoulos, Christos Diou · 2020
Recent years have seen an increase in data availability and computational power, which has led to superior performance in training deep learning models for image classification. In many real-world use-cases, however, the training datasets come with noisy labels that have been automatically generated for a small subset of the available data. In this paper we explore strategies for maintaining classification performance when the labels become noisy and sparse. In particular, we evaluate the effectiveness of combining PENCIL, a framework for correcting noisy labels during training and AMDIM, a self-supervised technique for learning good data representations from unlabeled data. We find that this combination is significantly more effective when dealing with sparse and noisy labels, compared to using either of these approaches alone.