Hardware Implementation for Multiple Activation Functions.

In a neural network, the activation function defines the output of that node. In this paper, we propose a novel hardware implementation for AI hardware accelerators to support three popularly used activation functions, including Hyperbolic Tangent, Sigmoid and ReLU. Based on this hardware implementation, the users can configure the activation functions of nodes at the execution time. Therefore, a higher flexibility for AI hardware computation can be provided. Implementation results show that the proposed hardware works well in practice.