ODDBALL 2.0: INTRODUCING VOLATILE STATISTICS TO STUDY LEARNING FROM UNEXPECTED OUTCOMES

The (auditory) oddball protocol is one of the most widely used tools in cognitive and systems neuroscience. Although the design reliably elicits mismatch responses towards oddball stimuli, a key limitation is that these responses do not drive learning, as stimulus statistics are stationary. This is at odds with the premise that unexpected outcomes should prompt learning to enable adaptive behaviour. To address this contradiction, we developed “oddball 2.0”. We started with a conventional passive oddball design with high- and low-frequency tones (standard, 90%; oddball, 10%; frequencies, 1kHz and 2kHz; tone duration, 0.5s; inter-tone-interval, 1s). Critically, we introduced a 5% chance of a state change after every stimulus. A state change reversed the mapping between frequency and oddball status. We hypothesized that after every state change participants would update their predictions to the new stimulus statistics. We showcase the appeal of our protocol by measuring pupil-linked arousal in human participants (N=20). We observed multiple pupil-based signatures of prediction updating: (i) larger tone-evoked pupil responses for state-dependent oddballs vs standards (p<0.001), (ii) exponentially decreasing pupil responses to standards after state changes (tau=0.86), and (iii) a linear relationship between pupil responses and the “change point probability” (p<0.001), derived from a Bayesian belief updating model. We conclude that phasic pupil-linked arousal supports the updating of predictions after unexpected outcomes. We advocate for the implementation of this novel oddball 2.0 protocol in study contexts ranging from animal models, to investigate underlying mechanisms, to clinical populations, to characterize aberrant prediction updating.