Metody uczenia sieci neuronowej Hopfielda

W artykule przedstawione zostały od strony teoretycznej i porównane od strony praktycznej różne metody uczenia sieci neuronowej Hopfielda. Oprócz znanej i powszechnie stosowanej reguły Hebba, przedstawione zostały modyfikacje tej metody. W celu porównania reguł uczenia sieci Hopfielda napisana została specjalna aplikacja, w której zaimplementowane zostały przedstawione w artykule metody. Regułą najlepiej rozpoznającą zapamiętane wzorce okazała się metoda pseudoinwersji

The Hopfield neural network can have many applications, such as approximation, compression, association, steering or patterns recognition. If the neural network is used for association, it is an associative memory. This task consists in original patterns recognition even when the Hopfield neural network is cued with distorted patterns. In this paper various learning methods for the Hopfield neural network are presented from the theoretical point of view and they are compared from the practical point of view. Besides the well known and generally used Hebb rule, there are presented its modifications as well. In order to compare the learning methods for the Hopfield neural network, a special application in which there are implemented the methods described in the paper is written. Section 2 contains the Hopfield neural network model, the Hopfield neural network definition and the neural network general schematic. There is also de-scribed the activation function used for testing the Hopfield neural network. Section 3 gives various Hopfield network learning rules, such as the original Hebb method, its modifications, the Oja rule and pseudoinversion rule. In Section 4 the testing process and its results are presented. The main task of this neural network is patterns recognition. The Hopfield neural network stored 10 patterns. Each of the stored patterns had 35 neurons. Then the neural network was cued with distorted patterns. The tests proved that the pseudoinversion rule recognized the patterns in the best way.