There proposal is actually that gradient descent is a bad representation of the NN learning behavior, and that instead this NFM tracks the “learning” behavior o

There proposal is actually that gradient descent is a bad representation of the NN learning behavior, and that instead this NFM tracks the “learning” behavior of the model better than simply its loss function over training.