Hardware Aware Training for Efficient Keyword Spotting - tinyML Research Symposium 2021

Explore cutting-edge research on hardware-aware training for efficient keyword spotting in this tinyML Research Symposium 2021 presentation. Delve into the development of new neural networks based on the Legendre Memory Unit (LMU) that achieve state-of-the-art accuracy with low parameter counts. Learn how these networks can run efficiently on standard hardware and custom-designed accelerator hardware, significantly improving power efficiency compared to general-purpose and specialized hardware. Discover the potential applications of this technology in mobile and edge devices, including phones, wearables, and cars. Gain insights into the high-level goals, LMU architecture, hardware-aware training techniques, accuracy vs. model size trade-offs, power modeling, and design space considerations for keyword spotting systems.

Introduction
High Level Goals
LMU
Hardware Aware Training
Accuracy vs Model Size
Power Modeling
Design Space
Comparisons
Questions
Sponsors