AI on a Microbudget - Methods of Machine Learning Miniaturization
summary
Shows concrete miniaturization recipes: reuse MobileNetV2 (3.5M parameters) via transfer learning to get a flower classifier with minutes of laptop training; distill OpenAI Whisper (1.55B, trained on ~700k hours) into Distil-Whisper and claim near-original English performance using ~2% of the original training data via knowledge distillation; apply model quantization to cut model size up to ~16× and run medium models (e.g., 15B params) in ~9 GB RAM. Accompanying GitHub repo supplies runnable notebooks for each method.
AI-written summary · about
chapters (5)
- 0:00Introduction to microbudget AI
- 2:58Transfer learning explained
- 8:46Model distillation techniques
- 15:30Quantization for smaller models
- 21:37Q&A and closing remarks
related talks
15mDesign Compact Deep Learning Models: Small is the New Big
31mKeynote: The next big thing in machine learning is tiny
39mCreating smaller, faster, production-worthy mobile machine learning models for Android
25mMachine Learning on Microcontrollers; Tiny, Useful and Fun
30mCompactifAI: Quantum-Inspired AI Model Compression
20mArchitecting AI: Efficient Models, Everywhere
46mSmall AI is the next Big Thing
16mExploring the world of Small Language Models