Generalization in Backpropagation Networks: An Empirical Study Using Image Data.

摘要

In this paper, we report the ability of trained multi-layer, first-order feedforward networks to generalize the perspective-invariant classification of image data. We introduce the method of interactive training which is useful when specific types of non-target images are important in the classification. Within interactive training, we show that the network can recognize perspective-independent images of one object and reject perspective-independent images of other objects in its training set, as well as reject a significant number of images of other objects on which it was not trained. However, the first-order network is still not capable of reliable one-class generalization. Therefore, we introduce a method for training a network with images on the boundary of the target class for one-class generalization. We report on the generalization ability of a second-order network trained with this method. We also describe the implementation of the feedforward network in a fixed-point hardware implementation that can process images at more than a billion connections per second. 16 refs., 9 figs.