Text this: Reducing Memory and Computational Cost for Deep Neural Network Training with Quantized Parameter Updates