Red neural artificial del sistema mesolímbico-cortical que simula el aprendizaje discriminativo y de inversión

Abstract

The present study develops a connectionist neural network with unsu pervised learning rules to simulate a discrimination task in a reduced number of time steps without previous training. The design of the network took into account some neurophysiological findings of dopaminergic mesolimbic system from structures like amygdala (AMG), orbitofrontal cortex (COF), ventral tegmental area (ATV) and nucleus accumbens (ACC). The proposed model generated similar responses to those from male rats during a discrimination and reversal learning tasks in a T maze, using sex as reward. In the activity of simulated structures different phenomena were found, like reinforcement preference and its reversal during reversal learning phase in ACC and ATV. It was also found an early encode in AMG, besides a retarded encoding and an increase in recruitment of neural nodes in COF during reversal learning. All output structures showed an expectancy activity before reinforcer delivery.