Per-Tensor Fixed-Point Quantization of the Back-Propagation Algorithm