Deep Learning in the world of little ponies
summary
Builds a My Little Pony image classifier using Keras convolutional nets and transfer learning, trained on roughly 100 images per class (three classes) at 150×150 resolution; the custom CNN contains about 3.5 million parameters and is trained in batches (e.g. 30 images, 10 epochs). Results show decent class scores but clear overfitting risk on a tiny dataset; also demonstrates a quick, accurate alternative via customvision.ai and notes use of an Azure GPU VM for reasonable training speed.
AI-written summary · about
chapters (16)
- 0:00Start
- 3:43What is an image?
- 6:44Classic ML vs. Deep Learning
- 9:34What is a Neuron?
- 11:17Layers in Neural Network
- 12:48Activation functions
- 14:41Backpropagation
- 23:52Networks: MNIST Classification
- 32:10CNN: Convolution Layer
- 33:40CNNs: Pooling layer
- 45:05Additional techniques: Data augmentation "Generate more training samples given a small dataset"
- 45:26Additional techniques: Dropout
- 46:01Transfer learning
- 47:01Pre-trained models: Feature extraction
- 48:55Pre-trained models: Fine-tuning
- 53:45Takeaways: Resources
related talks
01h 02mDeep Learning in the world of little ponies
53mDeep Learning in the world of little ponies
47mDeep Learning in Computer Vision
02h 26mPractical Deep Learning for Data Scientists
47mDeep Learning with PyTorch
04mUsing Computer Vision to Understand Big Visual Data
32mYou Can Do Deep Learning!
51mA Practical Introduction to Convnets for Image Classification