In this thesis, three bio-inspired algorithms viz. genetic algorithm, particle swarm optimizer (PSO) and grey wolf optimizer (GWO) are used to optimally determine the architecture of a convolutional neural network (CNN) that is used to classify handwritten numbers. The CNN is a class of deep feed-forward network, which have seen major success in the field of visual image analysis. During training, a good CNN architecture is capable of extracting complex features from the given training data; however, at present, there is no standard way to determine the architecture of a CNN. Domain knowledge and human expertise are required in order to design a CNN architecture. Typically architectures are created by experimenting and modifying a few existing networks.
The bio-inspired algorithms determine the exact architecture of a CNN by evolving the various hyperparameters of the architecture for a given application. The proposed method was tested on the MNIST dataset, which is a large database of handwritten digits that is commonly used in many machine-learning models. The experiment was carried out on an Amazon Web Services (AWS) GPU instance, which helped to speed up the experiment time. The performance of all three algorithms was comparatively studied. The results show that the bio-inspired algorithms are capable of generating successful CNN architectures. The proposed method performs the entire process of architecture generation without any human intervention.