Deep Learning in the world of little ponies
summary
Shows how to go from ~150–300 labeled pony images (about 220 train, 80 validation) to a usable classifier using Custom Vision Service for quick models and Keras + TensorFlow for custom CNNs, training on GPU to reach ~0.92 training accuracy with lower validation accuracy, and addresses overfitting with data augmentation and dropout. Demonstrates transfer learning with VGG16/ImageNet, exporting compact models and a Core ML mobile app, and exposing a REST prediction endpoint.
AI-written summary · about
chapters (18)
- 0:00Start
- 4:22Publish model
- 12:53What is an image?
- 13:57What is a Neuron?
- 15:27Layers in Neural Network
- 16:18Activation functions
- 18:24Backpropagation
- 20:44Putting things into practice
- 24:44Densely Connected Neural Networks: MNIST Classification
- 28:24CNN: Convolution Layer
- 29:48CNN: Pooling layer
- 31:20Information distillation
- 42:12Typical issues in the world of DL
- 43:09Overfitting
- 44:18Data augmentation or "squeeze more training samples from your small dataset"
- 45:46Transfer learning
- 48:40Demo: Use pre-trained models for classification task
- 52:20Takeaways: Resources
related talks
01h 02mDeep Learning in the world of little ponies
55mDeep Learning in the world of little ponies
40mDeep Learning on iOS
47mDeep Learning in Computer Vision
02h 26mPractical Deep Learning for Data Scientists
49mKeynote: From Zero to ML on Google Cloud Platform
01h 01mA practical guide to deep learning
40mMachines are Learning