|GALIMBERTI ANDREA||Cycle: XXXV |
Section: Computer Science and Engineering
Tutor: AMIGONI FRANCESCO
Advisor: FORNACIARI WILLIAM Major Research topic
:Energy-performance design of machine learning algorithms for IoT devicesAbstract:
The PhD research will focus on the hardware design of IoT platforms to integrate cognitive and intelligent behaviors while meeting the strict low power requirements. Voice assistants represent a critical example of the proposed research. Such devices implement voice recognition tasks on top of battery-powered IoT devices and represent an active research topic.
The Ph.D. research will be organized in three parts:
1. A complete review of the state-of-the-art represents the first step of the Ph.D. investigation. First, we will scrutinize the current hardware solutions to design cognitive and intelligent IoT platforms. Particular emphasis will be devoted to the analysis of the available open-hardware platforms, that potentially offer a way to integrate the PhD research and to deliver a working prototype.
2. The design and implementation of hardware solutions to support machine learning-based tasks represent the core of the Ph.D. proposal. The research is not solely focused on the design of hardware accelerators, but, in contrast, it considers the possibility to enhance the general purpose CPU and the ISA to offer machine learning support. The actual contribution will be the design of a low-power general-purpose CPU supporting machine learning tasks.
3. All the research will be integrated into the Lightweight Application-Specific Modular Processor (LAMP) platform, an open-hardware System-on-Chip implementing a RISC-V ISA for the main processing unit, developed at Politecnico di Milano. The PhD research will leverage the robust hardware design skills of the LAMP team to deliver a working prototype of the investigated methodologies in the 3-year time frame.
The research output is twofold:
1. A novel hardware design framework to integrate machine learning-based tasks into an open-hardware platform for IoT is offered. The proposed methodology will represent a solid standpoint to support future investigations in this research field.
2. The hardware-side investigation of state-of-the-art machine learning-based algorithms represents a critical step to improve the understanding of their energy-performance optimization. In particular, energy and performance aspects must be evaluated to deliver cognitive and intelligent smart IoT devices under severe low power requirements.