Bayesian surprise attracts human attention

By Laurent Itti, Pierre Baldi. 2005

1,751 citations recorded by the source corpus.

We propose a formal Bayesian definition of surprise to capture subjective aspects of sensory information. Surprise measures how data affects an observer, in terms of differences between posterior and prior beliefs about the world. Only data observations which substantially affect the observer’s beliefs yield surprise, irrespectively of how rare or informative in Shannon’s sense these observations are. We test the framework by quantifying the extent to which humans may orient attention and gaze towards surprising events or items while watching television. To this end, we implement a simple computational model where a low-level, sensory form of surprise is computed by simple simulated early visual neurons. Bayesian surprise is a strong attractor of human attention, with 72% of all gaze shifts directed towards locations more surprising than the average, a figure rising to 84% when focusing the analysis onto regions simultaneously selected by all observers. The proposed theory of surprise is applicable across different spatio-temporal scales, modalities, and levels of abstraction.

Subfields: Cognitive Neuroscience, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Cognitive Neuroscience

In Computer Vision and Pattern Recognition

Citations in the corpus

Automated Curriculum Learning for Neural Networks (2017, methodology): “…used as reward signals to encourage exploration, including compression progress (Schmidhuber, 1991), information acquisition (Storck et al., 1995), Bayesian surprise (Itti and Baldi, 2009), prediction gain (Bellemare et al., 2016) and variational information maximisation (Houthooft et al., 2016).”
Information seeking, curiosity and attention: computational and neural mechanisms (2013, positive): “2011) and may be quantified using Bayesian inference as a difference between a prior and posterior world model (Itti and Baldi 2009) or as a high prediction error for high-confidence states (Oudeyer and Kaplan 2007).”
Variational Information Maximisation for Intrinsically Motivated Reinforcement Learning (2015, background): “…or Bayesian surprise, which uses the KL divergence to measure the change in an agents internal belief after the observation of new data [8, 22]; measures based on prediction errors of future states such predicted L 1 change, predicted mode change or probability gain [16], or salient…”
Surprise-Based Intrinsic Motivation for Deep Reinforcement Learning (2017, background): “The Bayesian surprise associated with a transition is the reduction in uncertainty over possibly dynamics models from observing it ([1],[8]): DKL (P (φ|ht, at, st+1)||P (φ|ht)) .”
Active inference and epistemic value (2015, background): “An alternative perspective on this epistemic quantity is afforded by Bayesian surprise; namely, the KL divergence between prior and posterior beliefs (Bruce & Tsotsos, 2009; Itti & Baldi, 2009).”
Using Human Gaze to Improve Robustness Against Irrelevant Objects in Robot Manipulation Tasks (2020, neutral): “…theﬁxationlocationsalongasubject’sscanpath[43],similarity (SIM)whichisthesumoftheminimumvaluesateachpoint inthedistributionsthatarescaledtosumtoone[42],andthe Kullback–Leiblerdivergence(KL)[44].”
Graph-Based Visual Saliency (2006, positive): “Recently, Bruce [5] and others [4] have hypothesized that fundamental quantities such as "self-information" and "surprise" are at the heart of saliency/attention.”
Motion Guided Spatial Attention for Video Captioning (2019, background): “Prior researches of human perception (Itti and Baldi 2005; Howard and Holcombe 2010) have shown that human attentions are also more likely to be drawn to the apparently changing regions of a video.”
Salient object detection: A survey (2014, background): “prise [35], image quality assessment [36], [37], scene typicality [38], [39], aesthetic [33], and attributes [40].”
On scientific understanding with artificial intelligence (2022, background): “Alternative intrinsic rewards for artificial agents are computational creativity [106, 107] and surprise [108].”
Computational visual attention systems and their cognitive foundations: A survey (2010, methodology): “units corresponding to neurons in the human brain, but it has to be noted that they are still a high abstraction from the processes in the brain. Examples of connectionist systems of visual attention are presented in Olshausen et al. [1993], Postma [1994], Tsotsos et al. [1995], Baluja and Pomerleau”
Curiosity-driven Exploration in Deep Reinforcement Learning via Bayesian Neural Networks (2016, background): “This principle can be traced back to the concepts of curiosity and surprise [16–18].”
Latent Space Autoregression for Novelty Detection (2018, background): “The latter is mathematically modeled in terms of low probability to occur under an expected model, or by lowering a variational free energy [15].”
State-of-the-Art in Visual Attention Modeling (2013, background): “For instance, it has been claimed that visual attention is attracted to the most informative [144], the most surprising scene regions [145], or those regions that maximize reward regarding a task [109].”