Benchmarking Self Supervised Representation Learning methods on Biological Datasets

Deep Learning, and especially Deep Supervised Learning has been the driving force of computer vision in the last decade. Although this approach is the most promising when applied to a big dataset of annotated images, as the labeled data decrease in quantity, also its performance decreases in quality. This is ever so true when applied to highly specialized domains, such as Biological Imaging. For this very reason, scientists have been developing the so-called self-supervised paradigm. This paradigm allows learning a model even with a small number of annotated data, without compromising the quality of generalization power of the learned model. In this work we explore different ways of performing self-supervised learning, exploiting the new Pytorch-based framework: PytorchLightning. We run multiple experiments on different biomedical datasets, on NVidia-powered GPU computing nodes. Our main goal is to benchmark different self-supervised learning approaches and find out which one works best in the biomedical domain. As a sideproduct of this work, we produce a reusable and scalable codebase, which will speed up research by removing completely the boilerplate, that allows for fast prototyping and experimenting.