Scaling up the evaluation of recurrent Neural Network Models for cognitive neuroscience

Details

Supervisors

Crevecoeur, Frédéric ; Barak, Omri

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

The brain, perhaps the most complex entity in the known universe, is intimately connected to us and yet, remains so mysterious. It gives rise to our minds and subjective experiences, though its inner workings bear little resemblance to our conscious perception. Instead, looking inside brains reveals a myriad of different cells, known as neurons, forming an immensely complex network. Neuroscientists have endeavored to unravel this mystery by exploring how the electrical activity within these neurons, relate to familiar phenomena such as body movement, thoughts, language, emotions, and more. Significant progress in Artificial Intelligence (AI) not only led to the development of advanced AI assistants like ChatGPT, but has also created new avenues to better understand the brain. This thesis investigates how AI methods can be employed to study the connection between neuronal activity and cognition. Specifically, the focus will be on Recurrent Neural Networks (RNNs), a class of models capable of solving cognitive tasks that have dynamic aspects. This thesis will delve into the analysis of task-optimized RNNs, showing how dynamic behavior can be produced by dynamic networks of neurons. Existing methods, predominantly qualitative in nature, face limitations in scalability and are often applicable only to simple tasks. To address these issues, we have built a high-throughput pipeline for training different RNN models on a wide range of tasks and comparing them to experimental datasets through a variety of analysis methods. In particular, it includes quantitative metrics to evaluate models in a systematic and scalable way. We think this constitutes an important step towards more integrated models of the brain.