Dynamic Optimal Training for Competitive Neural Networks

This paper introduces an unsupervised learning algorithm for optimal training of competitive neural networks. The learning rule of this algorithm is derived from the minimization of a new objective criterion using the gradient descent technique. Its learning rate and competition difficulty are dynamically adjusted throughout iterations. Numerical results that illustrate the performance of this algorithm in unsupervised pattern classification and image compression are also presented, discussed, and compared to those provided by other well-known algorithms for several examples of real test data.