Visual search as active inference


Date
Sep 14, 2020 6:00 PM
Event
IWAI*2020
Location
Ghent (Belgium), gone virtual
  • What:: talk @ 1st International Workshop on Active Inference (IWAI*2020)

  • Who:: Emmanuel Daucé and Laurent Perrinet

  • Where: Ghent (Belgium), gone virtual, see https://laurentperrinet.github.io/talk/2020-09-14-iwai

  • When: 14/09/2020, time: 12:20:00-12:40:00

  • What:

    • Slides @ https://laurentperrinet.github.io/2020-09-14_IWAI
    • Code for slides @ https://github.com/laurentperrinet/2020-09-14_IWAI/
    • Abstract: Visual search is an essential cognitive ability, offering a prototypical control problem to be addressed with Active Inference. Under a Naive Bayes assumption, the maximization of the information gain objective is consistent with the separation of the visual sensory flow in two independent pathways, namely the “What” and the “Where” pathways. On the “What” side, the processing of the central part of the visual field (the fovea) provides the current interpretation of the scene, here the category of the target. On the “Where” side, the processing of the full visual field (at lower resolution) is expected to provide hints about future central foveal processing given the potential realization of saccadic movements. A map of the classification accuracies, as obtained by such counterfactual saccades, defines a utility function on the motor space, whose maximal argument prescribes the next saccade. The comparison of the foveal and the peripheral predictions finally forms an estimate of the future information gain, providing a simple and resource-efficient way to implement information gain seeking policies in active vision. This dual-pathway information processing framework is found efficient on a synthetic visual search task and we show here quantitatively the role of the precision encoded within the accuracy map. More importantly, it is expected to draw connections toward a more general actor-critic principle in action selection, with the accuracy of the central processing taking the role of a value (or intrinsic reward) of the previous saccade.