Hands-On Machine Learning with Scikit-Learn and PyTorch

Chapter 11. Training Deep Neural Networks

In Chapter 10 you built, trained, and fine-tuned several artificial neural networks using PyTorch. But they were shallow nets with just a few hidden layers. What if you need to tackle a complex problem, such as detecting hundreds of types of objects in high-resolution images? You may need to train a much deeper ANN, perhaps with dozens or even hundreds of layers, each containing hundreds of neurons, linked by hundreds of thousands of connections. Training a deep neural network isn’t a walk in the park. Here are some of the problems you could run into:

• You may be faced with the problem of gradients growing ever smaller or larger when flowing backward through the DNN during training. Both of these problems make lower layers very hard to train.

• You might not have enough training data for such a large network, or it might be too costly to label.

• Training may be extremely slow.

• A model with millions of parameters risks severely overfitting the training set, especially if there are not enough training instances or if they are too noisy.

In this chapter we will go through each of these problems and present various techniques to solve them. We will start by exploring the vanishing and exploding gradients problems and some of their most popular solutions, including smart weight initialization, better activation functions, batch-norm, layer-norm, and gradient clipping. Next, we will look at transfer learning and unsupervised pretraining, ...