Appropriateness of numbers of receptive fields in convolutional neural networks based on classifying CIFAR-10 and EEACL26 datasets

Abstract

An open question is studied of how many receptive fields (filters) a convolutional layer of a convolutional neural network should have. The goal is to find a rule for choosing the most appropriate numbers of filters. The benchmark datasets are principally diverse CIFAR-10 and EEACL26 for using a common network architecture with three convolutional layers whose numbers of filters are changeable. Heterogeneity and sensitiveness of CIFAR-10 with infiniteness and scalability of EEACL26 are believed to be relevant enough for generalization and spreading of the appropriateness of filters’ numbers. The appropriateness rule is drawn from top accuracies obtained on 10 × 20 × 21 parallelepipeds for three image sizes. They show, knowing that the number of filters of the first convolutional layer should be set greater for the more complex dataset, the rest of appropriate numbers of filters are set at integers which are multiples of that number. The multipliers make a sequence similar to a progression, e. g., it may be 1, 3, 9, 15 or 1, 2, 8, 16, etc. With only those multipliers, such a rule-of-progression does not give the number of filters for the first convolutional layer.