Research database

The current progress in Artificial Intelligence and Deep Learning technologies has an unsustainable environmental, social and economic cost. We identify the primary sources of this cost in a poor scaling of the computational cost of the training algorithms, on the one hand, and in a non-optimized structure of the hardware to execute these algorithms, on the other. In this project we propose a radically new framework for Deep Learning applications in physical substrates, that is based on the principles of hardware-software co-design and Green AI, making efficiency a key factor of both algorithms and their hardware implementations. On the hardware side, we embrace the Neuromorphic Computing paradigm, designing low-cost and versatile substrates based on Nanowire Networks for physical implementation of unconventional computing paradigms. On the software side, we focus on a physics-informed design of Reservoir Computing neural networks, using complex dynamical systems characterizations to enable the development of extremely efficient Deep Learning algorithms. The project will ultimately develop a Neural Nanowire Network (NNWN), a neuromorphic computing system capable of learning efficiently from temporal data. We will develop its mathematical description, software implementation, and physical fabrication. The NNWN system will be demonstrated, both in silico and on chip, in real-world edge-computing applications, validating its accuracy-efficiency tradeoff advantage in comparison to state-of-the-art solutions.