Quantified accelerated method to speed up inference?

[pajuhaan]

Hi, nice to see something in the C like this. this type of code can definitely bring the execution of AI models closer to hardware like DSPs or microcontrollers.
Seeing this project made me want to ask you this question; is it possible to speed up the inference layer by accepting some errors in weights and calculations, using simple shift registers (with pre-calculated shifts) and adders instead of floating/fixed point multiplier units? This approach could be applied at both the code and hardware levels. Does this idea make sense?
It's a doc about it:
Quantified Accelerated Artificial Neural Network Neurons - PJ.pdf