Democratizing Deep Learning methods by means of AutoML tools

Resumo

Artificial Intelligence, and specifically Deep Learning, methods have become the state-of-the-art approach to deal with Computer Vision problems in almost any field. This growth is due to the large amount of images that is produced in a daily basis, the increment of calculation capacity thanks to the development of specific hardware, and the open-source nature of the tools that allow us to build Deep Learning models. Despite their success, Deep Learning methods have several drawbacks. This technology might be difficult to use and requires some experience and technical knowledge. In addition, there is not always an algorithm, or library, that produces the best results for all the situations; hence, it is necessary to know and try different alternatives. Moreover, Deep Learning methods require a large number of labeled images to produce accurate models. This might be a challenge in contexts like biomedicine where it is difficult to acquire large enough datasets of images, and the annotation of those images require expert knowledge. Finally, once Deep Learning models are built, they should be able to generalize to contexts that were unforeseen during their construction. However, in many cases, models do not work properly when they are used with images from a domain, or style, different to the one used for training those models --- this is known as the domain shift problem. The goal of our work has been to tackle the aforementioned problems by means of techniques and tools that are user-friendly. First of all, we have developed tools that allow users to create accurate Deep Learning models for image classification and object detection tasks. To this aim, we have applied AutoML techniques that automatically search the best model for a given dataset of images. Moreover, we have developed a method to apply data augmentation to several Computer Vision problems. Such a method has been implemented in a tool that allows users to generate large enough datasets of images to feed Deep Learning models in several Computer Vision tasks such as image and video classification, object detection or semantic segmentation. In addition to tools that simplify the construction of Deep Learning models, we have developed a tool that tackles the domain shift problem by means of unpaired image-to-image translation methods and style transfer techniques. It is worth noting that we have not only developed methods and tools from a theoretical point of view, but all the knowledge acquired during the development of those tools has been applied to deal with actual biomedical problems such as spheroid segmentation, the classification and segmentation of motility images, or the diagnosis of retinal diseases from fundus images. Finally, the experience provided by tackling actual problems has served to improve the developed tools.