Exploring Techniques to Build Efficient and Robust TinyML Deployments

Explore techniques for building efficient and robust TinyML deployments in this 57-minute tinyML Talk. Delve into the challenges of edge deployment for deep learning applications, including privacy concerns, low power requirements, and robustness against out-of-distribution data. Learn about trade-offs between power and performance in supervised learning scenarios, and discover a dynamic fixed-point quantization scheme suitable for edge deployment with limited calibration data. Examine the compute resource trade-offs in quantization, such as memory and cycles. Gain insights into edge deployment architecture that utilizes deep learning methods to handle out-of-distribution data caused by sensor degradation and alien operating conditions. Topics covered include TinyML vs CloudAI, data considerations, transfer learning, keyword spotting, quantization techniques, architecture tweaking, and out-of-distribution detection.

Introduction
Agenda
TinyML vs CloudAI
Data
Transfer Learning
General Blueprint
Keyword Spotting
Quantization
Tweaking the architecture
Quantization and accuracy
Out of distribution detection
Out of distribution vs in distribution