Title:Nish: A Novel Negative Stimulated Hybrid Activation Function

Abstract:Activation functions play a crucial role in the performance and stability of neural networks. In this study, we propose a novel non-monotonic activation function is called Negative Stimulated Hybrid Activation Function (Nish). It behaves like a Rectified Linear Unit (ReLU) function for values greater than zero, and a sinus-sigmoidal function for values less than zero. The proposed function incorporates the sigmoid and sine wave, allowing new dynamics over traditional ReLU activations. We evaluate robustness of the Nish for different combinations of well-established architectures as well as recently proposed activation functions using on various well-known benchmarks. The results indicate that the accuracy rates obtained by the proposed activation function are slightly higher than those obtained using the set of weights calculated by Mish activation.