Beginning Anomaly Detection Using

downsampling.

The 

pooling layer helps reduce the size of the data to allow for easier computation. 

Additionally, it can help with pattern identification because the maximum value in each 

region is selected, allowing for the patterns to stand out more.

The 

dropout layer is next. Dropout is a regularization technique where a proportion 

(this is a parameter passed in) of randomly selected nodes are “dropped,” or ignored 

during the training process.

Flatten is a layer where the entire input is squashed into one dimension. Assume 

that you are trying to flatten a 3x3 image, like Figure 

3-41

.

Figure 3-40.  What a max pooling operation looks like on a 4x4 image

Chapter 3   IntroduCtIon to deep LearnIng

---

102

The 

dense layer is simply a layer of regular nodes similar to those in the artificial 

neural network example. They perform in the same way, but in this case the number of 

nodes varies from 128 in the first dense layer and to 10 in the second dense layer. The 

activation function also changes, from ‘relu’, or