Abstract

L. Itti, P. F. Baldi, A Surprising Theory of Attention, In: Proc. IEEE Workshop on Applied Imagery and Pattern Recognition (AIPR), Oct 2004. (Cited by 15)

Abstract: The concept of information is central to science, technology and biological function. Shannon's theory of information, although eminently successful for the development of modern computer and telecommunication technologies, does not capture subjective and semantic aspects of information that are not related to transmission but rather to observer expectations. Here we propose a subjective definition of information we call surprise, to quantify how data affects a (natural or artificial) observer, by measuring the difference between prior and posterior distributions of observer belief over families of models for the data. Surprise requires averaging over the space of models, contrasting with Shannon entropy which averages over data. We argue that biological sensory neurons signal quantities closer to surprise than Shannon information. To test this, we build a biologically-plausible computational model of early vision and bottom-up attention, which topographically computes low-level visual surprise. The model outperforms Shannon information-based models in predicting eye movement recordings of four human observers watching 50 complex video stimuli, including television broadcast. The resulting surprise theory of attention and subjective information foraging is applicable across different modalities, datatypes and abstraction levels.

Themes: Model of Bottom-Up Saliency-Based Visual Attention, Computational Modeling, Computer Vision, Bayesian Theory of Surprise