Automatic Design of Convolutional Neural Networks using Grammatical Evolution
Ricardo H. R. Lima, Aurora Pozo, Roberto Santana · 2019
The use of Convolutional Neural Networks (CNNs) has been demonstrated to be a solid approach for solving many machine learning problems, such as image classification and natural language processing tasks. Usual CNN architectures are composed of many convolutions, pooling and fully connected layers, from which the networks also learn a suitable representation for the data being processed. The manual design of CNNs is a complex task due to the high number of possible parameter configurations. Recent studies about automatic design of CNNs have shown positive results. Since it can be expressed as a hyperparameter optimization problem, in this study we propose to explore the design of CNN architectures through the use of Grammatical Evolution (GE). GE is a grammar based approach where a grammar is used to define the CNN components and structural rules. We performed a set of experiments using two well-known image classification datasets, the MNIST and CIFAR-10. The obtained results show that the presented approach achieved competitive results, while maintaining relatively small architectures, when compared with similar state-of-the-art approaches.