Practical Machine Learning with H2O

Chapter 8. Deep Learning (Neural Nets)

Deep learning is the new and trendy name for neural networks, and it sure is trendy! But deservedly so, as it is behind some of the most spectacular advances in AI and machine learning at the moment. Deep learning algorithms tend to be the best performers at problems that humans find easy yet (other) machine learning approaches find difficult, such as pattern recognition. Theoretically they can solve any problem, but they have their downsides too: they can be slow, they are black boxes and cannot explain their thinking, and they struggle a bit with categorical inputs. If your problem is to take a company’s annual transactions and calculate how much tax is owed, a neural net is not the right choice.

If you’ve used another library for neural nets or deep learning, one complaint you won’t have about H2O’s implementation is ease of use. As we saw back in Chapter 1, it takes care of most of the details for you, and you can get good results with a one-liner. Yes, there are still a huge number of parameters to tune but, as we will see in this chapter, the majority of them never need to be touched.

As in the other chapters, we will take look at how they work, but only the parts you need to understand to effectively tune them, then we will go through the parameters, and then dive into using deep learning on each of our data sets, first with defaults, then going through the tuning process.

However, a few special points to note. First, the use of deep ...