Design Compact Deep Learning Models: Small is the New Big
summary
Claims that automated model optimization—using pruning and quantization, reinforcement learning and genetic search plus compiler-level graph tuning—cuts model size from 80MB to 2MB (or 13MB to 0.5MB) and latency from about 1s to 100 ms per frame, enabling vision inference on Raspberry Pi/ARM microchips and even $5-class hardware instead of server GPUs. Benchmarks show preserved accuracy while reducing operations and memory, and toolchains produce deployable inference engines for edge autonomous perception.
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:33Embedded AI 101
- 2:06Deep learning models are growing rapidly
- 3:05These demands force AI to the cloud
- 3:50Edge computing challenges
- 4:27Time to deploy AI on edge devices
- 5:01Edge computing solutions: small is the new big
- 5:31Designing compact deep learning models
- 6:40Where does this fit in an ML/AI workflow?
- 7:25Levels of optimization
- 8:29Types of optimization
- 10:07Optimization benchmarks - computer vision
- 11:15Accelerating autonomous perception
- 11:53AI on low cost, low power chips
- 13:07Bringing AI to daily life
related talks
48mThe Power of Edge AI: Revolutionizing Future of Real-Time, Autonomous Processing
30mAI on a Microbudget - Methods of Machine Learning Miniaturization
33mHigh Performance Deep Learning on Edge Devices With
20mArchitecting AI: Efficient Models, Everywhere
30mCompactifAI: Quantum-Inspired AI Model Compression
39mCreating smaller, faster, production-worthy mobile machine learning models for Android
21mThings I learned running neural networks on microcontrollers
15mQuantization at the Edge: Making a 4GB Model Run on 1GB RAM