Deep neural networks yield the state of the art results in many computer vision and human machine interface tasks such as object recognition, speech recognition etc. Since, these networks are computationally expensive, customized accelerators are designed for achieving the required performance at lower cost and power. One of the key building blocks of these neural networks is non-linear activation function such as sigmoid, hyperbolic tangent (tanh), and ReLU. A low complexity accurate hardware implementation of the activation function is required to meet the performance and area targets of the neural network accelerators. This paper presents an implementation of tanh function using the Catmull-Rom spline interpolation. State of the art results are achieved using this method with comparatively smaller logic area.