Low-Power Hardware Implementation for Parametric Rectified Linear Unit Function

摘要

PReLU (Parametric ReLU) function is an advanced ReLU (Rectified Linear Unit) for improving the accuracy of the neural network. However, up to now, there is no special attention to the hardware design of PReLU function. Based on the characteristics of PReLU function, in this paper, we propose a dedicated multiplier for PReLU function. With this dedicated multiplier, we develop a low-power PReLU hardware implementation. Experimental results show that the proposed approach can significantly reduce power consumption with small accuracy loss.