TensorFlow Machine Learning Projects

Defining the loss function

The loss function for capsule networks is composed of two parts:

• Margin loss: Margin loss is exactly the same as what's used in Support Vector Machines (SVM). Effectively, we want the digit capsule to have an instantiation vector for class k, but only if the label is class k. For all other classes, we don't require any instantiation parameters. For each digit capsule k, we define separate loss as : 

  

If an image belongs to class k, then  else 0.  are the other two parameters.  is used for stability when initial learning ...