Automated Design of Feature-Aware Approximate Computing Units for Neural Networks

With the exponential growth of deep neural network model sizes, traditional computing architectures have faced severe energy-efficiency challenges. In this study, we propose a feature-aware approximation computing unit automation design method for neural network acceleration, that achieves synergistic optimization of accuracy and energy efficiency through the combination of dynamic quantization and configurable hardware units. First, based on the statistical characteristics of the activation values of the convolutional layers, a dynamic bit-width allocation strategy is proposed to adaptively adjust the computational accuracy of each layer by combining the sparsity characteristics of the ReLU activation function. Second, an approximation unit library is constructed, which contains existing common approximation adders, and improvements are made on the original basis to provide selectable adders for the automated design platform for training. Finally, a cross-level joint automated selection platform is constructed to achieve comprehensive optimization from algorithmic models to hardware logic. Experiments show that in mainstream models, such as VGG16 and ResNet18, the classification accuracy (Top-5) loss of this method on the ImageNet subset is less than 2%, whereas the energy-efficiency ratio reaches 2.3 TOPS/W, which is a 2.1 times improvement over the traditional exact computing architecture. This study provides an efficient hardware solution for large-scale neural network deployment.