TinyML for All: Full-stack Optimization for Diverse Edge AI Platforms

Explore full-stack optimization techniques for diverse edge AI platforms in this tinyML Summit 2022 conference talk. Delve into the challenges of deploying TinyML on resource-constrained devices and discover innovative solutions to improve neural network efficiency. Learn about model compression, neural architecture rebalancing, and new design primitives that address the mismatch between AI models and hardware. Gain insights into deploying real-world AI applications on tiny microcontroller units (MCUs) despite limited memory and compute power. Examine topics such as patch-based inference, MCUNet-v2, Once-for-All Network, NetAug for TinyML, TinyTL, and full-stack LIDAR and point cloud processing. Understand the fundamental problems in TinyML and explore how OmniML's approach of compressing models before training enables TinyML for various vision tasks.

Intro
TinyML is about Constraints
Everything Together: Real-world Al on Tiny MCUS
Brief History of MCUNets
Opportunity in Fundamental ML Algorithms
New Problem: Imbalanced Memory Distribution of CNNS
Solving the Imbalance with Patch-based Inference
MCUNet-v2 Takeaways
Once-for-All Network
Problem in Training for Tiny Models
NetAug for TinyML
Problem: Training Memory is much larger
TinyTL: Up to 6.5x Memory Saving without Accuracy Loss
Differentiable Augmentation
TinyML for LIDAR & Point Cloud
Full Stack LIDAR & Point Cloud Processing
Takeaways: Coming Back to MCUNets
Fundamental Problems in TinyML
OmniML "Compress" the Model Before Training
OmniML: Enable TinyML for All Vision Tasks
Founding Team