Putting AI on a Diet: TinyML and Efficient Deep Learning

Explore cutting-edge techniques for efficient deep learning and TinyML in this plenary talk from tinyML Asia 2021. Discover how to put AI on a diet as MIT EECS Assistant Professor Song Han presents innovative approaches to model compression, neural architecture search, and new design primitives. Learn about MCUNet, which enables ImageNet-scale inference on micro-controllers with only 1MB of Flash, and the Once-for-All Network, an elastic neural architecture search method adaptable to various hardware constraints. Gain insights into advanced primitives for video understanding and point cloud recognition, including award-winning solutions from low-power computer vision challenges. Understand how these TinyML techniques can make AI greener, faster, and more accessible, addressing the global silicon shortage and enabling practical deployment of AI applications across various domains.

Intro
Today's Al is too Big
Deep Compression
Pruning & Sparsity
Once-for-All Network: Roofline Analysis
OFA Designs Light-weight Model, Bring Alto Mobile Devices
NAAS: Neural Accelerator Architecture Search
Application Specific Optimizations
TinyML for Video Recognition
TinyML for Point Cloud & LIDAR Processing
SpAtten: Sparse Attention Accelerator
TinyML for Natural Language Processing
Tiny Transfer Learning