Principal Component Networks: Parameter Reduction Early in Training