Modelling Individual Humans via a Secondary Task Transfer Learning Method

Abstract

This thesis presents a novel approach towards modelling individual human behaviour on tasks with insufficient data via transfer learning. In most cases, deep neural networks (DNNs) require a great deal of data to train and adapt towards a particular problem. But there exist different tasks in which we do not have sufficient data available to train DNNs from scratch. There are approaches, such as zero-shot and few-shot learning, that can produce high quality DNNs with smaller amounts of data. However, these approaches still assume a large source dataset or a large secondary dataset to guide the transfer of knowledge from a source task to the target task. These are not assumptions that hold true when our goal is to model individual humans, who tend to produce much less data. In this work we present a novel transfer learning method for producing a DNN for modelling the behaviour of a specific individual on an unseen target task, by leveraging a small dataset produced by that same individual on a secondary task. We make use of a specialized transfer learning representation and Monte Carlo Tree Search (MCTS). We demonstrate that our approach outperforms standard transfer learning approaches and other optimization methods on two different human modelling domains.