Efeitos da variação da intensidade de aprendizado lateral no modelo de rede neural multicolunar do trion

Abstract

The Trion structured multicolumnar neural net model is caracterized by having two types of connections, internal and lateral, between its basic units, the trions, and by the fact that its states evolve synchronously. Each trion represents an idealized group of neurons, and a set of trions represents an idealized cortical column. Previous results have shown that each set of connections between the trions takes to a fixed repertoire of cyclic firing patterns (MP's). The evolution of each trion inside the MP's is influenced by its two previous firing states and also by its neighbors' two previous states. The model also allows learning by means of a hebbian algorithm, that effects small changes in the connections. In this work, we tried to verify the influence of various lateral learning intensities in the Cycle Probability (CP, the robustness of an MP) of different MP's with also different sets of connections. The results obtained showed that increases in lateral learning intensity, with some exceptions, take to an increase in the CP of an MP, followed by stabilization and a plateau. Many MP's had a CP nearing 1.0. Also were observed scapes with multiple minimuns and maximuns for the CP depending on the slat's values until the plateau was reached. In a general manner, the CP for the out of phase MP's were higher than the one for the in phase MP's.