Investigating the neural computations underlying the learning of a delay response task

Abstract

The behaviour of experimental animals reflects their physical and cognitive state. Animal models are a fundamental tool and resource to study such states. When analysing behavioural studies, different learning patterns can be distinguished: a gradual improvement or a sudden understanding. The former is a progressive method used for developing a new behaviour by dividing it into several stages. In addition to gradual improvement, learning also occurs by abrupt understanding, also known as aha moment, which is defined as a moment of abrupt insight or discovery. Lately, recent development of deep neural networks has had a remarkable impact on animal research. One strategy that has emerged as a promising tool for investigating the behaviour of animals performing a task is to study recurrent neural networks (RNNs) whose connection weights have been optimized to perform the same tasks as trained animals. In this work we have created simulated networks that emulate the learning processes in animals. Specifically, we have trained Long Short-Term Memory (LSTM) networks, which are a special type of RNN, with a shaping protocol on a Delayed Response (DR) task, that is a typical approach for studying mice behaviour. For this purpose, we have used Reinforcement learning (RL), which concerns goal-oriented algorithms. In order to analyse both mice and RNNs behaviour patterns, we have focused on the aha moment and compared their behaviours. We have complemented the study with an exploration of the effect of shaping in RNNs training.