Modeling Natural Action Selection

Neurophysiological evidence due to Schall,Newsome and others indicates that decision proc-esses in certain cortical areas (e.g. FEF and LIP)involve the integration of noisy evidence. Withinthis paradigm, we ask which neuronal architecturesand parameter values would allow an animal tomake the fastest and most accurate decisions. Sinceevolutionary pressure promotes such optimality(e.g. in prey capture and predator avoidance), it isplausible that biological decision networks realiseor approximate optimal performance. We considera simple decision model proposed by Usher &McClelland consisting of two populations of neu-rons integrating evidence in support of two alterna-tives, and we analyze the dynamics of this model.We show that in order to implement the optimaldecision algorithm (sequential probability ratiotest) the linearised network must satisfy the follow-ing two constraints: (i) it must accumulate the dif-ference between evidence in support of each alter-native, as would be implemented by mutual inhibi-tion between the populations; and (ii) the strengthof mutual inhibition must be equal to the leak ofactivity from each population.