Low-Bitwidth Generalized Feed-Forward Neural Network Digital Predistortion Based on Joint Optimization of Pruning and Quantization

This letter presents a novel double-stage quantization technique, jointly optimized with two pruning strategies, for lightweight generalized feed-forward neural network (GFNN)-based digital predistortion of 6G wideband signals. The method effectively reduces deep network size and converts floating-point (FP) operations into integer ones, which achieves substantial reductions in parameter count and memory footprint (MF). The proposed double-stage quantization adaptively adjusts quantization thresholds per training batch, improving overall quantization precision. It successfully lowers the bitwidths of activations/weights while maintaining a normalized root-mean-square error (NMSE) close to the full-precision baseline. Measurements using 200-MHz wideband signals validate that a 74% reduction in MF compared to the augmented real-valued time-delay neural network (ARVTDNN) model, without noticeable degradation in linearization performance.