Once the data has been labeled, we  need to adopt the prediction layer. Let's look at the following flow chart:

If we consider the image-classifier example used in the previous section, we label the image and depict different images as belonging to different classes.  

All of this information is fed to the neural network, which passes through the prediction layer using softmax, and this outputs the probabilities for each class. 

The neural network will map the output as 65% chances if it being class 0, 15% chance of the image being class 1, and so on. 

During the process of training, we calculate the difference of the real value and the predicted value. This information is used during backpropagation to vary the weights to obtain better predictions. 

The next prediction for the same example will be better and more accurate in nature. 

With the localization problem, the only difference is the structure of the data. Here's how we define our data:

We have P_c, followed by the properties of the bounding box and the class number, which looks identical to the classifier problem.

It's obviously no surprise that the prediction layer has to produce the same structure as the data. Due to this, the P_c value is a bit different because it's a probability:

To resolve this, values greater than 0.5 are treated as 1 and those less than 0.5 are treated as 0.

The prediction layer contains the bounding box properties and then the class numbers. We calculate the difference between the values. 

Consider the 0 example; we calculate the difference and then use it as feedback for backpropagation to change the weight to reduce the value of the difference between the predicted and real values.