An unsupervised learning algorithm for character recognition
C.K. Lee, Pui Fai Sum, K.S. Tan · 2003
An unsupervised learning algorithm is presented for character recognition based on the electrostatics phenomenon. It illustrates the application of a law of physics for learning. The authors treat the normalized pattern vectors as the position vectors of positive charges, and the normalized weight vectors as the position vectors of negative charges. The positive charges are fixed in location, and the negative charges move freely in the space. A step of movement of the negative charge indicates the change of the corresponding weight vector. Hence, based on the inverse square law, it is possible to evaluate each of the forces acting on a single charge. The direction of the change of a weight vector is along the direction of the resultant force. It was found that this network can be self organized to give various responses to different input patterns. Compared with other competitive learning algorithms, this algorithm can avoid the limitation due to differential initial settings.>