Attrattori dinamici nelle reti neurali ricorrenti

In recent neuroscientific literature, task representation has received a lot of attention. A major open question is how the architecture of clusters of neurons relates to the cognitive processes they execute and to the neurons’ functional specialization. Artificial neural networks are a powerful, widely employed tool to model and therefore comprehend the activity of these neural clusters. By training artificial networks to perform the same tasks of biological ones and looking at their structure and dynamics, significant insight can be gained into task representation by biological brains. This thesis focuses specifically on recurrent neural networks (RNNs), characterized by connections between layers which are shared over time and trained with the backpropagation-through-time method. The RNN at hand is trained to perform a simple cognitive task (“Go”), and the resulting response is analyzed both by examining the input and output signals, as well as the network parameters, and through the lens of a dynamical systems approach.