AutoPNN will implements a holistic cross-layer approach which aims in minimizing, in an automated manner, the hardware-requirements of the printed ML classifier at all possible levels in order to exploit the full potential of approximate computing and maximize the obtained hardware gains. To achieve this, AutoPNN applies a hardware-aware Neural Architectures Search (NAS) followed by neural minimization (software approximation) and an approximate hardware architecture search. At the software level, AutoPNN will produce a simplified PNN with the minimum possible hardware overheads. Hardware-driven software approximations can deliver significant hardware savings since they are able to remove coarse hardware structures at the expense of only a little accuracy loss.