Neural signatures of surprise in auditory learning
Introduction: Auditory mismatch responses (aMMRs) are neural signatures of learning evident when animals hear a surprising sound. Various mathematical models have been advanced to predict the amplitude, latency, and location of aMMRs.
Method: Using magnetoencephalography (MEG), I played auditory tones to participants using the Roving Oddball paradigm. I contrasted the spatiotemporal pattern of aMMRs and of gross brain activity (as measured by Global Field Power [GFP]) in different age groups, including in an understudied group aged 10-16. I used various mathematical models including the Bayesian Hierarchical Gaussian Filter (HGF) - and a variant of this that I modified to incorporate physical stimulus characteristics - to quantify the neural surprise likely to be evoked by auditory tones, examining whether this predicted brain activity.
Results: I found that aMMR amplitudes increased on trials where greater surprise was predicted. With age, anticipatory activity increased and aMMR amplitudes decreased. Unexpectedly, this decrease with age was not monotonic and older individuals’ aMMRs occurred at a lower latency than younger individuals’ aMMRs. The HGF predicted the magnitude of surprise better than traditional models did, especially on trials when the model predicted greater surprise. Incorporating physical pitch characteristics further improved model predictions. Unexpectedly, more repetitions of a tone were not always associated with less surprise and incorporating the precision of predictions and re-training the model continuously were not unambiguously found to improve predictions.
Discussion: My results broadly support the utility of Bayesian accounts of learning, with older individuals displays less evidence of neural surprise and greater anticipatory activity in relation to predictable stimulus features. However, the pattern of aMMRs found also suggests that brains may not implement Bayesian learning over purely categorical features, but also use continuous stimulus information, and that the role of precision is more nuanced. Some of the results with respect to age are arguably discordant with Bayesian predictions and more consistent with the developmental emergence of structure learning based on stimulus saliency.