Shallow Convolutional Neural Networks for Pattern Recognition Problems

Abstract

Paper describes an investigation of possible usage of shallow (limited by few layers only) convolutional neural networks to solve famous pattern classification problems. Brazilian coffee scenes, SAT-4/SAT-6, MNIST, UC Merced Land Use and CIFAR datasets were tested. It is shown that shallow convolution neural networks with partial training may be effective enough to produce the result close to state-of-the-art deep networks but also limitations are found.