Multiplierless MLP Using Successive Vector Approximation in Post-Training Quantization

Edge computing and IoT devices may have energy constraints that make it challenging to execute complex algorithms. In this paper, we propose the use of Sums of signed Powers of Two (SoPoT) as a quantization method for MultiLayer Perceptron (MLP) Neural Networks (NN) to reduce the computational burden. The so-called Matching Pursuits with Generalized Bit Planes (MPGBP) algorithm efficiently quan-tizes the coefficients of the MLP into SoPoT. It provides a clear tradeoff between the number of Signed Powers of Two (SPT) and the approximation quality, which is advantageous for computation under limited power availability. We evaluate the quantization impact in a detection problem for Multiple-Input Multiple-Output (MIMO) communication, where the Bit Error Rate (BER) for the detection using the SoPoT quantization and infinite precision are compared. The results show that the proposition reduces the model’s computational burden at the expense of performance losses, resulting in a tradeoff between computational cost and performance.