Matrix-Vectorized Canonical Signed Digit Quantized Neural Networks for Efficient Forward Pass Simulation

Quantized neural networks (QNNs) reduce computational cost but typically rely on multiplications, which limits their efficiency. Canonical Signed Digit (CSD) representation eliminates the need for multipliers through signed shifts. We present a matrix-vectorized Python framework for efficient batch forward-pass evaluation of CSD-QNNs, achieving near-lossless accuracy (absolute errors 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-5</sup>, MSE 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-10</sup>) and robustness across 8 - 16 bit quantization, while highlighting hardware potential for shift-and-add implementations in FPGA/ASIC deployments.