True 4-Bit Quantized Convolutional Neural Network Training on CPU: Achieving Full-Precision Parity

Low-precision neural network training has emerged as a promising direction for reducing computational costs and democratizing access to deep learning research. However, existing 4-bit quantization …