Neural networks were introduced in the chapter on Apache Spark MLlib. This chapter builds upon this understanding by introducing Deep Learning, which uses deep neural networks. These are neural networks that are feature-rich and contain extra hidden layers so that their ability to extract data features is increased.

These networks are generally feed-forward networks, where the feature characteristics are inputs to the input layer neurons. These neurons then fire and spread the activation through the hidden layer neurons to an output layer, which should present the feature label values.

Errors in the output are then propagated back through the network (at least in back propagation), adjusting the neuron connection weight matrices so that classification errors are reduced during training.

The following diagram is a good illustration:

The previous example diagram, described in the H2O booklet at https://leanpub.com/deeplearning, shows a deep learning network with four input neurons to the left, two hidden layers in the middle, and two output neurons. The arrows show both the connections between neurons and the direction that activation takes through the network.

These networks are feature-rich because they provide the following options:

• Multiple training algorithms
• Automated network configuration
• The ability to configure many options
• Structure
• Hidden layer structure
• Training
• Learning rate, annealing, and momentum

So, after giving this brief introduction to deep learning, it is now time to look at some of the sample Scala-based code. H2O provides a great deal of functionality; the classes that are needed to build and run the network have been developed for you. You just need to do the following:

• Prepare the data and parameters
• Create and train the model
• Validate the model with a second Dataset
• Score the validation Dataset output

When scoring your model, you must hope for a high value in percentage terms. Your model must be able to accurately predict and classify your data.