Evaluating the Effect of User-Given Guiding Attention on the Learning Process

Abstract

Most current supervised learning systems require large quantities of labelled data, limiting their applicability in domains where labelled data is scarce and hard to obtain. We introduce a novel approach for incorporating additional, user-given areas of interest during training by which the learning process can be guided. The provided guiding attention is incorporated in the training phase as a form of data augmentation, which ensures that input dimensions do not vary between train and test/deployment time, when no guiding attention is present. We evaluate this approach by extending the CIFAR-10 dataset with prototypical information and ascertain, that our approach reduces the required amount of samples by up to 44.89%, when combined with traditional data augmentation techniques. This would enable the use of learning systems in parts of manufacturing such as commissioning, where additional samples are scarce and costly to obtain while providing guiding attention is a matter of seconds.