Orientation-dependent effects in neural networks

Abstract

When human observers decide whether disoriented patterns are normal or reflected (mirror images), response time increases with orientational disparity from the upright. This orientation-dependent effect was obtained with a 4-layer neural network and a back-propagation algorithm. In the learning phase of Experiment 1, the network learned to classify normal-reflected patterns that could appear at different orientations on an input grid. In the test phase, the same patterns were presented at new orientations. Fewer trials were required to re-learn the classification when the patterns were closer to the upright. In Experiment 2, the same network was presented with new patterns. Fewer trials were again required to learn the classification when the patterns were closer to the upright.