Opening Keynote
summary
A live demo shows a Python bird-or-not classifier implemented in under 50 lines using a Keras-style library and ImageNet-pretrained weights, illustrating transfer learning from 1.2M ImageNet images. The talk explains convolutional filters, max‑pooling and 9 anchors per spatial position for region proposals, then describes Faster R‑CNN (two‑stage) versus SSD (single‑stage), noting Faster R‑CNN runs around 5–6 fps on GPU. An open‑source project called Luminol packages training, eval, dataset tooling and a web demo.
AI-written summary · about
chapters (23)
- 0:00Start
- 5:19Machine Learning Approach
- 6:14Supervised Learning
- 6:26Introduction to Neural Networks
- 7:30Convolutional Neural Networks
- 8:41What Is a Convolutional Filter
- 11:07Activation Map
- 14:34How Do We Know How Many Filters in each Layer
- 16:18Result of a Convolutional Layer
- 21:01The Activation Map
- 22:16Imagenet
- 24:56Geoff Hinton
- 26:39Object Detection
- 29:55Why Do I Want To Detect Objects
- 30:53Pinterest
- 34:06Two-Stage Object Detector
- 35:17Activation Maps
- 37:14Region Proposal
- 37:57Reference Box
- 42:50Classification Layer
- 43:22Summary
- 47:59Training on Google Cloud Ml
- 53:21Lumi Server Web
related talks
02h 26mPractical Deep Learning for Data Scientists
23mBirding with Python and Machine Learning
01h 27mHands-On Computer Vision with PyTorch
49mKeynote 4: Here Come The Robots - Python and Machine Learning
01h 39mGoogle Colaboratory and Python: Deep Learning basics for CV tasks
15mMy first steps on Deep Learning with Keras
43m(Deep) Learn You a Neural Net for Great Good!
01h 55mGoogle Colaboratory and Python: Deep Learning basics for CV tasks