Designing convolutional neural network architecture by the firefly algorithm

This paper presents firefly algorithm framework for designing convolutional neural network architecture. Convolutional neural networks can be classified as a special category of deep neural networks that in most cases consist of several convolution, fully connected (dense) and pooling layers. Wide set of image classification tasks and problems from the computer vision domain were successfully tackled by convolutional neural networks. One of the most challenging tasks from this domain is to find the convolutional neural network architecture that obtains the best performance for the specific application. The values of network's hyper-parameters have significant influence on the overall network performance. Research shown in this paper deals with convolutional neural network hyper-parameters optimization that define the network's architecture and structure. The hyper-parameters that were taken into account for this research include the number of convolutional and dense layers, the number of kernels per layer and the kernel size. We performed hyper-parameters optimization by the well-known firefly algorithm that belongs to the group of swarm intelligence metaheuristis. Solution's quality, robustness and performance of our proposed framework was tested against the MNIST dataset. Obtained empirical results showed that the proposed framework achieves promising performance in this domain.