Beginning Anomaly Detection Using

Max 

pooling is where the input data is scanned by a filter, which in this case is a 2x2 filter, and 

the maximum value in the 2x2 region of the image is chosen to be the value in the new 

n-dimensional image. If the 

stride length is not given, by default Keras chooses the pool 

size. The stride length is how far the filter should shift, and it plays a role in determining 

the feature map size. In this case, since the stride length is 2 and the pooling filter size is 

also 2x2, the dimensions of the input data are reduced in half.

Assume that the 4x4 image in Figure 

3-40

 is the input to a max pooling layer with 

pool size of 2x2.

Figure 3-39.  Once the filter reaches this value here, the convolution operation 

ceases, outputting a feature map to the next layer

Chapter 3   IntroduCtIon to deep LearnIng

---

101

Since the pool size is 2x2 and the stride length is also 2 in this case (no parameter 

was provided for stride length), the pooling layer happens to split the entire image into 

regions of 2x2 pooling filters.

If the stride length was 1, then you would have a situation similar to the convolution 

example you saw earlier, and the dimensions of the feature map would be 4-2+1 = 3x3. 

This process of pooling can also be referred to as