A data-centric approach for improving ambiguous labels with combined semi-supervised classification and clustering

Life is not black and white -- Combining Semi-Supervised Learning with fuzzy labels

“…An assumption for the reason for this performance was that fuzzy images are not easily classifiable into the existing classes and thus confuse the algorithm. The fuzzy labels can be incorporate in the training procedure with for example S2C2 which is an extension to most existing SSL algorithms [4]. The second issue is that most algorithms aim at good hard classifications as output for the neural network.…”

“…We use the Mice Bone dataset proposed in [4] and show examples in Figure 1d. This dataset consists of gray-scale images of collagen fibers of mice bone.…”

“…The classification problem has three classes: similar, dissimilar and not relevant fiber structures. We follow [4,10] and use only consistent images for training and validation and enforce a class balance in this data. We call this consistent and balanced training data seed and use the rest of the data as unlabeled data.…”

“…One of the biggest issues with deep learning is the required amount of labeled data for training. Thus, many Semi-and Self-Supervised algorithms have been proposed which can decrease the required labeled data by using additional unlabeled data [1,2,3,4,5,6]. These methods aim for the clear vision that we feed the complete data into the network and additionally provide only few labeled samples per class.…”

“…For example, the cross bread of two different dogs can not be classified as one of its parents. These two issues have been summarized before as fuzzy labels [10,4] and have been discussed as problematic for many Semi-Supervised Learning algorithms. The issue of fuzzy labels is common in many life science [11,12,13] or medical [14] datasets.…”