Tensorflow, deep learning and modern convolutional neural nets, without a PhD
summary
Builds a plane detector starting from a 20x20 classifier trained in TensorFlow + ML Engine using TensorBoard, the Estimator API and the Dataset API; converts 256x256 images into ~5,000 20x20 crops, runs hyperparameter tuning: 50 trials, 10 parallel and hits ~99.6% test accuracy but poor real-world precision. Switches dense nets to convolutional stacks (benchmarks SqueezeNet vs Darknet), finds global average pooling breaks localization, switches to reshape+dense and a YOLO-style grid (8x8→16x16) with improved bounding-box assignment and usable detection.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:05Why I dont read papers
- 1:52The reality of deep learning
- 3:47What is a neural network
- 5:28Tensorflow code
- 9:40Rayleigh function
- 15:18Convolutional Neural Networks
- 21:55Hyperparameter tuning
- 27:04Estimator
- 35:57Inception paper
- 38:38Global average pooling
- 39:35Squeezenet
- 40:45Darknet
- 42:26Results
- 44:39How does it work
- 47:36The results
- 49:08Demo
- 50:42Conclusion
related talks
55mA Tour of Deep Learning With C++
01h 33mIntroduction to Convolutional Neural Networks using TensorFlow and Keras
01h 07mDeep Learning with TensorFlow
38mDeep Learning for detection on a phone
01h 06mDeep Learning with TensorFlow
47mDeep Learning in Computer Vision
01h 29mComputer Vision: An (Un?)Expected Journey, with Keras and Tensorflow
01h 01mA practical guide to deep learning