Text this: Mutual Information Based Learning Rate Decay for Stochastic Gradient Descent Training of Deep Neural Networks