Biologically plausible learning algorithms for recurrent neural networks

The very first works that gave light to the field of Artificial Intelligence were heavily influenced by the study of the brain, and carried out by collaborative efforts of computer- and neuro-scientists. Progressively, the models and methods proposed have distanced themselves from this approach, towards ones more oriented to informatics. In the last years, a resurgence of interest in works that bridge the two fields has taken action, thanks also to development of tools that allows to investigate the brain with a extraordinary level of detail. Biological mechanisms act as a natural source of inspiration for innovative techniques applicable to artificial networks. The project of this internship positions itself in this trend. The objective is to implement a training procedure for recurrent neural networks starting from considerations on the cerebellum. A form of Stochastic Gradient Descent has been proposed to be performed by this structure, involved in the coordination of motor actions. Activity of neurons is perturbed in a stochastic manner that results in a change of the global error. Synaptic plasticity is carried out integrating information about perturbations and error evolution. Two training algorithms for recurrent neural networks are taken as a base for investigation on the temporal dynamics and learning in this special class of networks, and are progressively modified to implement the mechanism proposed for cerebellar learning.