Lightweight CNN Frameworks and their Optimization using Evolutionary Algorithms

This paper presents novel frameworks for efficient and lightweight convolutional neural networks that are suitable for use in embedded system applications. Population-based metaheuristic approaches including the genetic algorithm, cuckoo search, the multifactorial evolutionary algorithm and a proposed hybrid approach are used to optimize their performance on image classification tasks. The methods utilize small population sizes without requiring weight-sharing or a surrogate function, and both binary and integer encoding methods are applied in the optimization. The results from these various strategies are evaluated using metrics of computation time and classification accuracy. The multifactorial-based approach is found to give the highest classification accuracy and it requires only a moderate evaluation time. Also, comparisons with prior approaches demonstrate that these methods show a favorable tradeoff between accuracy and computational cost.