|DISABATO SIMONE||Cycle: XXXIV |
Section: Computer Science and Engineering
Tutor: GATTI NICOLA
Advisor: ROVERI MANUEL Major Research topic
:Deep and Wide Tiny Machine LearningAbstract:
In the last decades and, in particular, in the last few years, Deep Learning (DL) solutions emerged as state of the art in several domains, e.g., image classification, object detection, speech translation and command identification, medical diagnoses, natural language processing, artificial players in games, and many others.
In the same period, following the massive spread of pervasive technologies such as Internet of Things (IoT) units, embedded systems, or Micro-Controller Units (MCUs) in various application scenarios (e.g., automotive, medical devices, and smart cities, to name a few), the need for intelligent processing mechanisms as close as possible to data generation emerged as well. The traditional paradigm of having a pervasive sensor (or pervasive network of sensors) that acquires data to be processed by a remote high-performance computer is overcome by real-time requirements and connectivity issues.
Nevertheless, the memory and computational requirements characterizing deep-learning models and algorithms are much larger than the corresponding abilities in memory and computation of embedded systems or IoT units, significantly limiting their application. The related literature in this field is highly fragmented, with several works aiming to reduce the complexity of deep learning solutions. However, only a few aim to deploy such DL algorithms on IoT units or even on MCUs. All these works fall under the umbrella of a novel research area, namely Tiny Machine Learning (TML), whose goal is to design machine and deep learning models and algorithms able to take into account the constraints on memory, computation, and also energy the embedded systems, the IoT, and the micro-controller units impose.
This work aims to introduce a methodology as well as algorithms and solutions to close the gap between the complexity of Deep Learning solutions and the capabilities of embedded, IoT, or micro-controller units.
Achieving this goal required operating at different levels. First, the methodology aims at proposing inference-based Deep Tiny Machine Learning solutions, i.e., DL algorithms that can run on tiny devices after their training has been carried out elsewhere. Second, the first approaches to on-device Deep Tiny Machine Learning training are proposed. Finally, the methodology encompasses Wide Deep TML solutions that distribute the DL processing on a network of embedded systems, IoT, and MCUs.
The methodology has been validated on available benchmarks and datasets to prove its effectiveness. Moreover, in a ``from the laboratory to the wild'' approach, the methodology has been validated in two different real-world scenarios, i.e., the detection of bird calls within audio waveforms in remote environments and the characterization and prediction of solar activity from solar magnetograms. Finally, a deep-learning-as-a-service approach to support privacy-preserving deep learning solutions (i.e., able to operate on encrypted data) has been proposed to deal with the need to acquire and process sensitive data on the Cloud.