Computational Modelling of Information Gathering

This thesis describes computational modelling of information gathering behaviour under active inference – a framework for describing Bayes optimal behaviour. Under active inference perception, attention and action all serve for same purpose: minimising variational free energy. Variational free energy is an upper bound on surprise and minimising it maximises an agent’s evidence for its survival. An agent achieves this by acquiring information (resolving uncertainty) about the hidden states of the world and uses the acquired information to act on the outcomes it prefers. In this work I placed special emphasis on the resolution of uncertainty about the states of the world. I first created a visual search task called scene construction task. In this task one needs to accumulate evidence for competing hypotheses (different visual scenes) through sequential sampling of a visual scene and categorising it once there is sufficient evidence. I showed that a computational agent attends to the most salient (epistemically valuable) locations in this task. In the next, this task was performed by healthy humans. Healthy people’s exploration strategies provided evidence for uncertainty driven exploration. I also showed how different exploratory behaviours can be characterised using canonical correlation analysis. In the next study I showed how exploration of a visual scene under different instructions could be explained by appealing to the computational mechanisms that may correspond to attention. This entailed manipulating the precision of task irrelevant cues and their hidden causes as a function of instructions. In the final work, I was interested in characterising impulsive behaviour using a patch leaving paradigm. By varying the parameters of the MDP model, I showed that there could be at least three distinct causes of impulsive behaviour, namely a lower depth of planning, a lower capacity to maintain and process information, and an increased perceived value of immediate rewards