Confronto di reti neurali ricorrenti addestrate con algoritmi standard e biologicamente plausibili

Recurrent neural networks (RNNs) are a versatile tool for neuroscience, as they provide a simplified model of biological neural circuits performing cognitive operations. The standard method to train RNNs is gradient descent, implemented via backpropagation. Albeit provably efficient, this learning method is fundamentally implausible from a biological point of view: animals do not learn from a large set of labeled examples, and biological neural circuits cannot realistically implement backpropagation. Therefore, more plausible learning algorithms have been proposed, which rest on random exploration and rewards. In this thesis, the student will simulate a simple RNN, and train it to perform a simple cognitive task via a biologically-plausible learning algorithm. The resulting RNN dynamics and performance will be compared with those of an RNN trained for the same task via standard gradient descent.