Improvement of supervised learning by linear mapping

In this paper, we propose a new supervised learning method to improve the learning speed for feedforward neural networks. Our method is different from traditional supervised learning methods that find a one-to-one mapping between the given-input pattern matrix and the desired output pattern matrix. Instead, it finds one of the one-to-many mappings between the input matrix and an intermediate output matrix, and transforms the intermediate output matrix to the desired output matrix in one step using linear mapping techniques. Learning is faster with our method because there exist many intermediate output matrices, and learning can stop whenever one such matrix is found. Our extensive experimental results show that our learning algorithm converges to within a reasonable range of error after a few training epochs, making it suitable for dynamic real-time applications in which the network may need to be re-trained periodically.