Beginning Anomaly Detection Using

convolution layer filters through the data and 

multiplies each of the values element-wise by the weights in the filter and sums them 

up to generate one value. In this case, it’s a 3x3 filter that slides over each of the pixels to 

generate a smaller layer called an 

activation map or feature map. This feature map then 

has another filter applied to it in the second convolutional layer to generate another, 

smaller feature map. The weights that are optimized during backpropagation are found 

in the filter. To get a better idea of this, let’s look at some examples of how this works.

Assume a 5x5 pixel picture like Figure 

3-34

.

Assume also that your kernel size (filter dimensions) is 2x2. Figure 

3-35

 shows how 

the convolutions would go.

Figure 3-34.  A 5x5 pixel picture, with 0 representing black pixels and 1 

representing white pixels

Chapter 3   IntroduCtIon to deep LearnIng

---

96

To begin with, you have a random set of weights for the 2x2