Application-aware optimization of Artificial Intelligence for deployment on resource constrained devices

Artificial intelligence (AI) is changing people's everyday life. AI techniques such as Deep Neural Networks (DNN) rely on heavy computational models, which are in principle designed to be executed on powerful HW platforms, such as desktop or server environments. However, the increasing need to apply such solutions in people's everyday life has encouraged the research for methods to allow their deployment on embedded, portable and stand-alone devices, such as mobile phones, which exhibit relatively low memory and computational resources. Such methods targets both the development of lightweight AI algorithms and their acceleration through dedicated HW. This thesis focuses on the development of lightweight AI solutions, with attention to deep neural networks, to facilitate their deployment on resource constrained devices. Focusing on the computer vision field, we show how putting together the self learning ability of deep neural networks with application-specific knowledge, in the form of feature engineering, it is possible to dramatically reduce the total memory and computational burden, thus allowing the deployment on edge devices. The proposed approach aims to be complementary to already existing application-independent network compression solutions. In this work three main DNN optimization goals have been considered: increasing speed and accuracy, allowing training at the edge, and allowing execution on a microcontroller. For each of these we deployed the resulting algorithm to the target embedded device and measured its performance.