A 28nm Energy-Area-Efficient Row-based pipelined Training Accelerator with Mixed FXP4/FP16 for On-Device Transfer Learning | IEEE Conference Publication | IEEE Xplore