Reduction of Neural Network Circuits by Constant and Nearly Constant Signal Propagation

This work focuses on optimizing circuits representing neural networks (NNs) in the form of and-inverter graphs (AIGs). The optimization is done by analyzing the training set of the neural network to find constant bit values at the primary inputs. The constant values are then propagated through the AIG, which results in removing unnecessary nodes. Furthermore, a trade-off between neural network accuracy and its reduction due to constant propagation is investigated by replacing with constants those inputs that are likely to be zero or one. The experimental results show a significant reduction in circuit size with negligible loss in accuracy.