In this chapter, we covered what deep neural networks are in more detail, particularly how to use them to train prediction models. Even though deep feedforward neural networks can seem quite complex, they can be broken down into a sequence of layers, each of which is fairly simple, with one set of inputs and one set of outputs, along with weights and biases to map between the two.

We have also seen the improvement in predictive performance possible using deep learning. In the use case example, using linear regression alone accounted for 23% of the variance in the testing data; however, by using a deep feedforward neural network, we were able to account for 35% of the variance in the year of song release. Although still far from perfect, it is a dramatic improvement over regression, and the low performance probably has more to do with lacking the data to explain year-to-year differences than the model itself (in other words, even with the best model achieving 99% variance accounted for is unlikely without more/better predictors). The next and final chapter will cover how to tune and optimize models, including how to address some common challenges such as missing data or poor model accuracy/performance.