Analysis of Feature Maps Selection in Supervised Learning Using Convolutional Neural Networks

Artificial neural networks have been widely used for machine learning tasks such as object recognition. Recent developments have made use of biologically inspired architectures, such as the Convolutional Neural Network. The nature of the Convolutional Neural Network is that each convolutional layer of the network contains a certain number of feature maps or kernels. The number of these used has historically been determined on an ad-hoc basis. We propose a theoretical method for determining the optimal number of feature maps using the dimensions of the feature map or convolutional kernel. We find that the empirical data suggests that our theoretical method works for extremely small receptive fields, but doesn’t generalize as clearly to all receptive field sizes. Furthermore, we note that architectures that are pyramidal rather than equally balanced tend to make better use of computational resources.