WE ALL know what surprise feels like, but a computer model has now defined the concept. It is the change in expectation caused by the arrival of new data, it says. The model uses an aptly named unit of measurement 鈥� the 鈥渨ow鈥�.
at the University of California, Irvine, and at the University of Southern California in Los Angeles devised the model while investigating human attention.
A dominant theory from the 1950s has it that the amount of attention we pay to an object or event is linked to the volume of information our brains need to form an understanding of it. For example, our attention should hover over intricate patterns longer than over a plain surface.
Advertisement
But this model did not consider that the majority of data is useless, says Itti, while only a little is of interest to you or might indicate a threat. Instead, he and Baldi reckon that we focus more on objects or movements that attract our attention by being surprising or unexpected. Surprise as they compute it may also explain what causes the 鈥渙rienting reflex鈥�, whereby our attention is caught by novel stimuli.
To test their hypothesis, the pair developed a computer model which simulated a population of visual neurons 鈥渨atching鈥� video clips, just as your brain would watch it through the eye鈥檚 retina. They used the model to analyse short video clips and mark which regions of the videos it considered the most surprising 鈥� which they rated in wows. 鈥淪omething that is very surprising has a high wow content,鈥� says Baldi.
When they showed the videos to human volunteers, their eye movements correlated with what the computer had rated as being worthy of attention.
鈥淲e found that human observers did indeed look at surprising things while watching TV, much more than they looked at information-rich objects,鈥� says Itti (Vision Research, ).
This study is a long-awaited 鈥渟atisfactory theoretical account鈥� for what holds our attention, says Aapo Hyv盲rinen, a computer scientist at the University of Helsinki in Finland. He adds that it formulates 鈥渁 Bayesian theory of surprise in which an event is surprising if it changes our beliefs鈥�.
Itti says the model could have wide-ranging applications. For example, it could be used to rank websites for interest, as those providing more original content would stand out, while spammers, copycats and aggregator sites may be classified as boring. It could also be used to design more eye-catching advertisements, he says.
He and Baldi are now carrying out experiments in monkeys to see if individual retinal neurons compute surprise in the way their model predicts.
