Adaptive Neural Networks for Agile TinyML - Keynote

Explore a groundbreaking approach to running neural networks on resource-constrained devices in this keynote address from the tinyML Summit 2021. Delve into the concept of adaptive neural networks that dynamically minimize memory and computational requirements during inference. Learn about the challenges facing tinyML, the basics of dynamic inference, and its relationship with hardware. Discover throttleable neural networks (TNN) and their intuitions, controller training techniques, and early results on hardware. Examine practical applications through object detection examples, metrics for agility, and development workflows. Gain insights into hardware accelerators and how this adaptive approach enables more flexible and efficient deployment of machine learning models on tiny devices.

Intro
Challenges for tinyML
Towards Dynamic and Adaptive
Basics of Dynamic Inference
Relationship with Hardware
Examples Modeling Approaches
Throttleable Neural Networks (TNN)
Basic Intuitions on TNN
Controller Training
Early Results on Hardware
Selecting The Best Utilization
Object Detection using TNN
Example Metrics for Agility
Example Development Workflow
Hardware Accelerators