Convolutional Neural Network Architecture Design by the Tree Growth Algorithm Framework
Ivana Strumberger, Eva Tuba, Nebojša Bačanin, Raka Jovanović, Milan Tuba · 2019
This paper presents tree growth algorithm framework for designing convolutional neural network architecture. Convolutional neural networks are a special class of deep neural networks that typically consist of several convolution, pooling and fully connected layers. Convolutional neural networks have proved to be a robust method for tackling various image classification tasks. One of the most important challenges from this domain is to find the network architecture that has the best performance for the specific application. The performance of the network depends on the set of hyper-parameter values such as the number of convolutional and dense layers, the number of kernels per layer and kernel size. Optimization of hyperparameters was performed by novel tree growth algorithm that belongs to the group of swarm intelligence metaheuristics. The robustness, performance and solutions quality of the proposed framework was validated against the well-known MNIST dataset. Conducted comparative analysis demonstrated that the proposed frameworks obtains promising results in this domain.