Lossless Compression of Neural Network Components in Low-Precision Formats