This is for optimizing inference. For consumer devices such as smart cameras, speech recognition on your phone, or robotics that's the part that matters. They

This is for optimizing inference. For consumer devices such as smart cameras, speech recognition on your phone, or robotics that's the part that matters. They have real speed improvements by using 1-bit quantitzed weights and then running operations using 32-bit or 64-bit bitfields which happen to use very few instructions/cycles on modern CPUS: https://github.com/allenai/XNOR-Net https://github.com/allenai/XNOR-Net