Understanding What Artificial Intelligence Actually Sees
summary
Claims that you can build a usable custom object detector with about 50 images per label by transfer learning from a pretrained vision model, trains on GPUs via Watson Machine Learning (training runs can take ~30 minutes) and exports formats for TensorFlow.js / TF Lite / CoreML. Demo moves inference into the browser with WebGL shaders so detections run in tens of milliseconds (vs 1s+ for a cloud API or slow plain JS), model files are ~20 MB, and the provided object-detection-react app loads a downloaded model and performs real-time bounding-box detection.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:04What is object detection?
- 0:29Classification
- 4:02Deep learning is a bunch of matrix math
- 7:00How do we train?
- 8:29Let's play a game!
- 10:28We can use this!
- 11:22Let's play another game!
- 15:39Miscommunication
related talks
32mReal-Time Object Detection with CoreML
41mAI in Web: TensorFlow js
26mMachine Learning and AI for Angular Developers
29mImage Processing in the Browser
55mCore ML: Vision
49mReal Time Face Detection through a Webcam Using AI
38mReal-world Machine Learning with TensorFlow and Cloud AI
15mInferenceJS: Real-time computer vision in your browser