On Yerkes-Dodson and naturalistic attention

When giving a talk I always try to follow the rule ‘first tell them something that they know, and then tell them something that they don’t know’ (for reasons discussed here).

For this reason I always start a talk on arousal and attention by talking about Yerkes-Dodson – the thing that most people know about arousal - that it's a U-curve:

There is in fact, though, surprisingly little evidence either for or against this model, as far as I can tell, in the human literature (although see here, here, here and here). And of course, what research there is has used trial-by-trial experimental paradigms rather than naturalistic, ‘real-world’ data. So how might we address this using ‘real-world’ data?

One experiment I’d love to do when I get a chance, that makes some nice predictions, is to collect a large body of ‘free-play- viewing data, using a head-mounted eyetracker. An eyetracker like this has two cameras on it – one recording the participant’s field of view (everything they can see) and the other recording where within the field of view they are looking.

Using off-the-shelf toolboxes you could then compute the saliency of the field of view data. And using ROC curves you could compute how predictive salience was of where the child was looking (in epochs, say of 5 minutes at a time). And, for bonus points, you could even hand-code semantic information (such as whether the child was looking at faces, which are semantically important but often quite low salience in naturalistic viewing data).

And your predictions? Nice and clean. Salience should be most predictive of location of gaze during hyper-arousal. During mid-level arousal and during hypo-arousal, salience is less predictive. And semantic information should be most predictive of location of gaze during mid-level arousal. During both hypo- and hyper-arousal, it should be less predictive. All nice and strong predictions arising from Yerkes-Dodson - and just waiting to be tested, should anyone get the chance!