Pattern classification using trainable logic networks

摘要

The author describes a new pattern classification algorithm whichhas the simplicity of the well-known multilinear classifier but iscapable of learning patterns through supervised training. This isachieved by replacing the discretely valued logic functions employed inthe conventional classifier with continuous extensions. The resultingdifferentiable relationship between network parameters and outputspermits the use of gradient descent methods to select optimal classifierparameters. This classifier can be implemented as a network whosestructure is well suited to highly parallel hardware implementation.Essentially, the same network can be used both to compute weightadjustments and perform classifications, so that the same hardware couldbe used for both rapid training and classification. The author hasapplied this classifier to a noisy parity detection problem. Theclassification error frequency obtained in this example comparesfavourably with the theoretical lower bound