All Times ET
Keywords: Self-supervised learning, robusteness
Although machine learning has made considerable progress in recent decades, most methods still rely on immense, high-quality datasets, which limits the areas of application. Self-supervised learning offers a promising solution to this difficulty by applying a pretext task based on self-generated pseudo-labels, which does require a high-level of semantic understanding of the data, enabling the network to learning meaningful visual representations and permitting the model to solve more complex downstream tasks. While self-supervision can be characterized by a reduced performance compared to supervised models, it improves robustness. In this paper we demonstrate that self-supervision is superior especially for small data sets and combinations of different data manipulations, but also delivers significantly better results for data loss and unbalanced data sets. For very small datasets, we can achieve an accuracy 12.2 percentage points higher than the standard approach. Furthermore, we investigate the influence of the pretext task complexity on the results as well as the combination of different data manipulations thoroughly. This enables us to generate a better understanding of the relationship between data, pretext and downstream task and the robustness of the model.