10 AIs or 10B AIs
Talkat Understanding Higher-Level Intelligence from AI, Psychology, and Neuroscience Perspectives · 2024-06-24..06-28·
summary
Proposes training from a single continuous video stream (concatenating ~10 walking‑tour videos of ~1h40 each, ~700 object classes per video) using a DINO variant called DoRA that adds a self‑supervised multi‑object tracker to produce masks; pretraining with guided future prediction (Kinetics) helps; continuous 24‑hour streams train with batch size 1 but require different optimizers—RMSProp over Adam—and less momentum; updating every step helps in‑stream while accumulating (≈16 chunks) improves off‑stream generalization; results approach ImageNet pretraining on several tasks.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:05Current AI models limits
- 4:03Embodied AI vision
- 8:01Relaxing the need for large diverse datasets
- 18:46Learning from video
- 26:53Continual learning
- 34:24Single-stream conclusions
- 35:58Q&A
related talks
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
01h 15mTrain Object Detection with Small Datasets
08mAgents and Robots: The Human Effort Putting AI Into Production
01h 30mDeep Learning Techniques for Monocular Depth Estimation
32mVideo Action Models
01h 29mHuman-like Visual Search Application with Small Data
59mDeep Neural Networks for Video Applications
47mBest Practices for Real-time Intelligent Video Analytics