Efficient presentations of learning samples to accelerate theconvergence of learning in multilayer perceptron

摘要

Four methods of presenting learning samples are proposed toincrease efficiency of learning in multilayer perceptrons. The methodsinvolve presenting samples selectively instead of randomly; typical andconfusing samples are selected and presented in systematic order. Themethods were simulated to examine their effectiveness in a simplethree-layer perceptron with two inputs and two outputs. All the methodsexcept the method of presenting typical samples alone turned out to besuperior to the conventional method of random presentation. The two bestmethods were to present typical samples in the first half period oflearning and confusing ones in the second half period of learning, andto present in turn both typical and confusing samples