Exploring the limits in learning capabilities of biologically plausible neural networks performing in a serial reversal task

C. J. MININNI; B SILVANO ZANUTTO

Congreso; Neuroscience 2017; 2017

SFN

It has been proposed that animals construct models of their environment to better adapt their behaviour and increase their chances of survival. To achieve this goal, animals have to search for patterns even when the adaptive value of a behavioural responses not only depends on the current scenario, but also in the history of events. In this regard, the Serial Reversal Task (SRT) is a behavioural paradigm in which two rules alternate over time, demanding the animal to keep track of previous events in order to maximize reward. Traditional neural network models cannot explain learning in the SRT because learning of one rule usually erases previously acquired information. The goal of this work is to find the essential properties required by stochastic spiking neural networks to solve a SRT. We found that the SRT cannot be solved if the integration of stimulus information and the decision process occur in the same neural population. This limitation is independent of the number of neurons, neuronal dynamics or plasticity rules, and stems from the fact that plasticity is locally computed at each synapsis, and that all stimulus/response pairings are equally reinforced. We propose a biologically plausible neural network model which solves the SRT, based on separating the function of integration of stimulus information from the function of response selection. The results shed new light about the functioning of decision-making brain structures like the prefrontal cortex, and highlight the importance of characterizing neural circuits based on their connectivity and the degree of plasticity modulation with the reward.