Recurrent Neural Network approaches to model primate behavior

Recurrent Neural Networks (RNNs) are increasingly used in theoretical neuroscience to model simple animal behavior. In this context, two conceptually different approaches have been taken: in the first approach, the RNN is trained on neural activity recordings of the animal during a task, to infer the structure of a recurrent circuit giving rise to the observed activity patterns; in the second approach, the RNN is trained on stimulus-response associations of a given task, to infer how a recurrent circuit might implement the related mapping; In this thesis, we start from previous work that trained RNNs on neural activity recordings of a monkey during a simple contextual decision making task, and train RNNs only on the corresponding stimulus-response association. The structure and dynamics of the trained RNNs are analyzed and compared with previous results,with a particular focus on how biological constraints (e.g., the implementation of Dale's principle and the necessity of distributed processing ) affect RNN behavior and its agreement with previous findings.