Programming Machine Learning

What You Just Learned

In this chapter, we used the Keras ML library to write a quick three-layered neural network, and train it on a simple dataset. That shallow network gave us decent results, but it seemed to have trouble dealing with fine details in the data.

We expected that a deeper network would do better, so we added a fourth layer to the network. The results were disappointing. While the deeper network proved 100% accurate on the training set, it got nowhere near that accuracy on the all-important validation set. In fact, the deeper network did worse on the validation set than the shallow three-layered network.

To recap, we made our neural network deeper, but that didn’t make it better. In fact, it seems that once our network gets deeper, ...