Embedding AI into Constrained Devices: A Multi-Faceted Approach

Thanks to their data-driven nature, machine and deep learning approaches have recently reached super-human performance, promoting the last artificial intelligence (AI) spring. Accordingly, the application of such techniques to the industrial world has vastly grown in popularity. However, the most common deep learning models, namely neural networks (NNs), are characterised by an intrinsic trade-off between performance and efficiency. Focusing on raw performance, recent efforts produced highly complex NN models made of several millions or even billions of parameters. This complexity hinders the application of AI into industrial devices and appliances characterised by limited computational capabilities and resources. Accordingly, in this thesis, we focus on the embedding AI into constrained devices problem, to which we refer to as the open research challenge of applying AI techniques to devices characterised by limited computational capabilities and resources. We tackle the embedding AI task reframing the problem from a NN efficientisation perspective, where the aim is the minimisation of the resource usage of NNs, either during their optimization process or their deployment phase. We propose a pioneer multi-faceted approach in which we consider both (i) the available efficientisation approaches – aiming at analysing and overcoming some of their limitations –, and (ii) to leverage Neuro-Symbolic integration (NeSy) mechanisms to tackle the efficientisation perspective. As a result of our twofold perspective, we shed new light on the NN efficientisation issue, highlighting the groundbreaking opportunities available leveraging NeSy systems.