Lecture - Theories of Deep Learning MT25, XI, Visualising the filters and response in a CNN
-
Transfer learning: Download a pre-existing net and then re-train the last layer, or last few layers
- Spectra of the net through the learning process
- Why does transfer learning work?
- When does transfer learning not work?
- Consider e.g. fine-tuning a network that learns to classify dogs vs cats in order to classify specific breeds: this is not a good idea since the model has explicitly learned to map out differences in dog and cat breeds
-
CNNs learn representations resembling wavelets / curvelets / ridgelets
- Attention sinks