Towards Adaptive and Robust Tiny Machine Learning on Multi-Core Embedded RISC-V MCUs


Manuele Rusci, Francesco Conti and Luca Benini

Presentation title

Towards Adaptive and Robust Tiny Machine Learning on Multi-Core Embedded RISC-V MCUs

Authors

Manuele Rusci, Francesco Conti and Luca Benini

Institution(s)

University of Bologna

Presentation type

Technical presentation

Abstract

Enriching sensor devices with on-board intelligence is a primary need to avoid network congestions caused by billions of connected devices that stream raw data. The on-device extraction of meaningful and compressed high-level information also preserves data privacy and reduces the power consumption of sensor nodes by duty-cycling the power-hungry transmission sub-systems. On the other side, the low available energy budget imposes the selection of low-energy computing units for local sensor processing. Thanks to the software flexibility and low-power consumption, MicroControllers (MCUs) result in the best fit for smart sensors. Unfortunately, bringing energy-efficient and robust AI pipelines, especially belonging to the Deep Learning field, - recently referred to as Tiny Machine Learning (TinyML) - on MCUs is an extremely challenging process given the scarcity of memory and computation resources featured by the target platforms.

This talk will present the recent methodologies to address this challenge for deeply embedded systems levering a multi-core RISC-V MCU engine. On the one side, we will discuss strategies and tools to compress and quantize Deep Learning inference models to fit the memory constraints of the target system. On the other side, we will present optimized software libraries and HW extensions to favor the efficient deployment of quantized DL models on battery-powered sensor nodes. The talk will show end-to-end applications solutions enabled thanks to the combination of the proposed HW/SW and tools. Lastly, we will focus on recent techniques to enable (Continual) learning adaptation on these ultra-low-power smart sensor devices to break the “inference-only” wall of present TinyML solutions.


Additional material

  • Presentation slides: [pdf]