Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators