Enhancing Convolutional Neural Network Performance through Optimized Sigmoid Activation Function Modeling
Ye-Rim Youn, Jin-Keun Hong · Asia-pacific Journal of Convergent Research Interchange · 2023
Artificial intelligence (AI) has recently attracted significant attention due to its involvement in phenomena like deep fakes, AI-generated artwork, and other prominent projects and collaborations.It is being used across various domains such as security and deep learning, necessitating highperformance hardware specifications centered around graphics processing units.The research methodology employed in this paper can be outlined as follows: the pursuit of optimal parameter values by manipulating the coefficients and constants within the function employed by the convolutional neural network, and fine-tuning the parameter values themselves.The outcome of this endeavor, involving the optimization of coefficients, constants, and parameter values in the convolutional neural network's functions, yields significant insights.For the original sigmoid formula, optimal parameter values are identified: a coefficient of 9 for the first "x," a coefficient of 7 for the second "x," and a constant term of 0.0001.Shifting attention to the modified Sigmoid function, the pursuit of optimal conditions for high accuracy and fast calculation time leads to the determination of specific values.Notably, the coefficient of the first "x" ("β") within the function is established as 30, while the coefficient of the second "x" ("δ") is set at 70.Additionally, the square value of 5 is identified as optimal, further underscoring the importance of achieving a balance between accuracy and computational efficiency.