A novel algorithm for identifying the optimal CNN Architectures regulated by Swarm Intelligence

Abstract

The Emergence of deep learning architectures in recent years paves a pathway for developing several real-time applications. Deep learning architecture selection is time-consuming since it takes hours to get trained, few architectures take days to get training. To address the pitfall in selecting the architectures, we developed a novel algorithm that returns an optimal architecture to the user with the help of swarm intelligence. The parameters required to run the algorithm have been obtained dynamically from the input samples, so the returned architecture is reliable. The algorithm is tested based on qualitative metrics such as training accuracy and testing accuracy, and the algorithm is evaluated based on quantitative metrics such as convergence time and execution time. The algorithm is robust against any dataset, and the architecture generated by the algorithm yields better results. • A novel algorithm for generating deep learning architectures by regulating swarm optimization technique is introduced. • A small set of input sample is enough to generate the optimal architecture from the available architectures. • Time taken for generating the optimal architecture is nominal. • We proved the accuracy of the dynamically generated architecture is better in terms of quantitative and qualitative metrics. • The algorithm can be used in any type of dataset to generate a better architecture by dynamically obtaining parameters from the input data. • This algorithm will pave a new pathway in the deep learning research domain.