Towards a Better Compromise Between Shallow and Deep CNN for Binary Classification Problems of Unstructured Data

Deep Neural Network is a large scale neural network. Deep Learning, refers to training very large Neural Networks in order to discover good representations, at multiple levels, with higher-level learned features. The rise of deep learning is especially due to the technological evolution and huge amounts of data. Since that, it becomes a powerful tool that everyone can use specifically on supervised learning, because it’s by far the dominant form of deep learning today. Many works based on Deep learning have already been proposed. However, these works have not given any explanation on the choice of the number of the network layers. This makes it difficult to decide on the appropriate deep of the network and its performances for a specific classification problem. In this paper the objective is threefold. The first objective was to study the effect of facial expressions on facial features deformations and its consequences on gender recognition. The second objective is to evaluate the use of Deep learning in the form of transfer learning for binary classification on small datasets (containing images with different Facial expressions). Our third goal is then to find a compromise between too much capacity and not enough capacity of the used deep Neural Network in order to don’t over fit nor under fit. Three different architectures were tested: a shallow convolutional neural network (CNN) with 6 layers, a deep CNN VGG16 (16 layers) and very deep CNN RESNET50 (50 Layers). Many conclusions have been drawn.