Development of the Learning Process for the Neuron and Neural Network for Pattern Recognition
Ramaz Khurodze · American Journal of Intelligent Systems · 2015
The goal of our work is development of a neuron and a neural network which recognize any object without mistakes. It worked out by learning method, with totality of simple (single-layer) neurons. The work deals with the formal neuron and neural network learning process through the realizations of a set of learning sets. At the same time, the features and the feature space used for the recognition process are evaluated. The feature space for all the types and realizations is of the same dimension and binary. The learning process is carried out by means of recognition procedures; this, in case of incorrect recognition, as much as possible, draws together the changes in neuron weighting coefficients as well as the threshold (structuring etalon descriptions) with the neural process of recognition. A set of realizations for the learning clusters of each pattern is used for the learning process. The learning algorithm comprises two stages. The first stage represents structuring etalon description of its own, the second – the correction of the received description in relation to other patterns of descriptions by using the same patterns of the learning set’s realizations. The correction of the results received in the recognition process is carried out by means of changing the weighting coefficients through using the award algorithm (procedure). In case of the incorrect recognition of some realization, it is presented to the neuron until we get the correct recognition through coefficients changing (error correction of mistakes) which may require neuron threshold changing.