Improving Neural Network Training in Low Dimensional Random Bases