S2C2 - An orthogonal method for Semi-Supervised Learning on ambiguous labels
Semi-Supervised Learning (SSL) can decrease the required amount of labeled image data and thus the cost for deep learning. Most SSL methods assume a clear distinction between classes, but class boundaries are often ambiguous in real-world datasets due to intra- or interobserver variability. This ambiguity of annotations must be addressed as it will otherwise limit the performance of SSL and deep learning in general due to inconsistent label information. We propose SemiSupervised Classification & Clustering (S2C2) which can extend many deep SSL algorithms. S2C2 automatically estimates the ambiguity of an image and applies the respective SSL algorithm as a classification to certainly labeled data while partitioning the ambiguous data into clusters of visual similar images. We show that S2C2 results in a 7.6% better F1-score for classifications and 7.9% lower inner distance of clusters on average across multiple SSL algorithms and datasets. Moreover, the output of S2C2 can be used to decrease the ambiguity of labels with the help of human experts. Overall, a combination of Semi-Supervised Learning with our method S2C2 leads to better handling of ambiguous labels and thus realworld datasets.
PDF Abstract
CIFAR-10
STL-10
