Investigation of the generalizing capabilities of convolutional neural networks in forming rotation-invariant attributes

This paper gives the results of a study of the possibilities of convolutional neural networks to generalize knowledge concerning primitive geometrical image transformations when solving pattern-recognition problems of handwritten numerals. Experiments were directed to the study of how the recognition of patterns in arbitrary orientations is affected by broadening the training sample with rotated images. Results are presented for convolutional neural networks of two architectures, showing that, to ensure rotation-invariant recognition, it is necessary for all classes of images in the entire range of rotations to be present in the training sample. (C) 2015 Optical Society of America.