Noniterative learning in perceptrons implemented by an ultrafast-learning character-recognition scheme
Chia-Lun John Hu · 2002
As we studied in the last five years, for an artificial perceptron consisting of hard-limited neurons, the connection matrix to meet a given input-output mapping can actually be obtained noniteratively in one step if the given mapping satisfies a certain PLI condition. Whenever the given mapping satisfies this condition, generally there exists infinitively many solutions for the connection matrix. One can then select an optimum solution such that in the recognition mode, the recognition of any untrained input vectors becomes optimally robust. The "learning" here (or the obtaining of the connection matrix from the given mapping) should be very fast because the learning process is noniterative and one-step. The recognition of untrained inputs here should be optimally robust because the optimum analysis here is independent of the learning method we use. This paper reports the theoretical analysis of this noniterative learning scheme and the design and the experiment of a practical ultrafast-learning, character-recognition scheme derived from this theory.