gaCNN: Composing CNNs and GAs to Build an Optimized Hybrid Classification Architecture

Convolutional Neural Networks (CNN) are considered the gold standard for Computer Vision Problems. However, finding the best architecture for CNN often requires handcrafted design and domain knowledge. On the other hand, Genetic Algorithms (GAs) have proven to be an efficient technique to optimize a wide range of problems. Thus, in this paper, we propose the gaCNN, a hybrid classification architecture composed of a CNN and a GA. The gaCNN utilizes heterogeneous activation functions to classify images, optimizing its hyperparameters and activation functions automatically, regardless of the analyzed dataset. The results show that gaCNN is able to identify good architectures. For the Fashion MNIST dataset, the gaCNN obtained, as best accuracy, 93.73%, better than 9 of 13 compared classifiers. For the MNIST Handwritten dataset, the gaCNN obtained, as best accuracy, 99.14%, better than 12 of 16 classifiers.