Sustainable Deep Learning for the Development of Autonomous Computational Intelligences

This project aims to develop science and technology to enable the implementation of Autonomous Computational Intelligences (ACIs). ACIs are software and hardware systems that, in combination with sensors and actuators, form intelligent entities capable of autonomously performing complex tasks in the physical or virtual world. They are capable of interacting with humans, machines and systems, such as traffic, to perform complex tasks, for instance, the delivery of products by an autonomous truck controlled by an ACI. Recently, thanks to the significant advances introduced by deep learning, neural networks have become essential towards the development of powerful ACIs. This deep learning revolution has been important to build trust in such a technology. However, the need to incorporate such knowledge in low computing power devices requires attention regarding processing speed and memory consumption. Moreover, acquiring and annotating problem-specific real-world datasets, especially in robotics and autonomous systems, is often a hard task requiring overwhelming human effort and, sometimes, specific expertise. Although the development of ACIs has advanced a lot with deep learning, it still requires cutting-edge computers to run models and process sensor data as well as expensive human annotation. Motivated by these aspects, this project proposes to investigate methods to reduce the computational and human burden of deep learning in order to enable the development of more powerful ACIs . This proposal aims to put forward a new and more sustainable pipeline for deep learning applications. The main scientific contribution will be an end-to-end pipeline that goes from data annotation and preprocessing to architecture task-wise optimization, using human knowledge to disambiguate only critical decisions.