Chapter 5

1. You do what you have to do... but ideally, you want to use it only once, in order to avoid bias in your choice of model.
2. It is the process of generating more data out of the initial dataset, to increase its size and improve the generalization of the network.
3. Not exactly: Keras replaces the original images with the ones coming from data augmentation.
4. In general, the dense layers tend to be the ones with the most parameters; in particular, the first one after the last convolutional layer is usually by far the biggest.
5. While the line of the training loss goes down with the number of epochs, the validation loss goes up.
6. Not always: you can use a strategy to first overfit the network (to properly learn the training dataset), and then to improve generalization and remove the overfitting.