Activation function comparison in neural-symbolic integration

• Main Authors: Mansor, Mohd. Asyraf, Sathasivam, Saratha
• Format: Conference Proceeding
• Language: English
• Subjects: C++ (programming language); Computer simulation; Logic programming; Neural networks; Simulation; Training; Visual programming languages
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/1.4954526

The activation function is a dynamic paradigm for doing logic programming in Hopfield neural network. In neural-symbolic integration, the activation function used to metamorphose the activation level of a unit (neuron) into an output signal. The proposed activation function is Bipolar sigmoid activation function. The main goal of this research is to compare and analyze the performance of proposed activation function paradigm with another sign function, namely McCulloch-Pitts function based on Wan Abdullah’s method. In this study, we evaluate experimentally the differences between both functions through computer simulations. Computer simulations are conducted to demonstrate the ability of Bipolar sigmoid function and McCulloch-Pitts function doing the logic programming in Hopfield neural network. Microsoft Visual C++ 2013 is used as a platform for training and testing. The performance of Bipolar sigmoid activation function and the McCulloch-Pitts function were discussed holistically by comparing the global minima ratio, Hamming distance and training or computation time. It was proven by computer simulations that the Bipolar sigmoid activation function has a better performance, provides good solutions and achieves an acceptable stability compared to the McCulloch-Pitts function.