Bayesian active-perception: an informatic viewpoint
thesisposted on 21.05.2010, 10:09 authored by David Heavens
Localisation of places, prey and predators are usually of critical behavioural importance to an organism’s survival. In this thesis, I conduct an investigation of localisation, considering specifically the inference of a target, event or the observer itself. I begin with an exploratory investigation into auditory localisation of a single sound source for a static (passive) observer. I evaluate the influence (sensitivity) of “cue” variables on localisation by the curvature of the location belief’s Kullback-Leibler divergence. More generally, from this I observed a symbol grounding problem – corresponding one location to a data sample due to multiple locations mapping onto a single observed value. I demonstrate how action can support the grounding of symbols by breaking such symmetries (inference confusions) that exist in passive localisation. By considering the breaking of these symmetries, I go on to develop an information measure that generally selects the best localising action. This is the action expected to give the “next best view” for the system, hence removing ambiguities and uncertainties in inference with the greatest efficiency. From these considerations, my main contribution is a general theoretical framework for selecting between actions during localisation and inference tasks according to an observer’s representation. I illustrate this framework by using it to select head casts in localising binaural level cues for sound source localisation. Further illustration is through a learning problem, where I evaluate learning performance during directed and undirected selection of actions. This demonstrates how directed action is important in symbol grounding of the latent state space to the observation space. Because of its generality, my Bayesian-active perception framework may be used to derive novel domain specific action-selection and learning algorithms that optimise inference. It may also provide a principled account for existing action-selection algorithms (for instance in robotics) and specific animal behaviours as special cases.