Prediction and surprise in human fear conditioning (Commentary on Boll et al.)


The amygdala has long been recognized as crucial for the processing of emotional information, especially fear and negative affect. In this article (Boll et al., 2012), the authors approach amygdala function in human fear conditioning with considerable subtlety. Using high-resolution functional magnetic resonance imaging, they track the updating of processing of both cues and outcomes as participants’ expectancies are first confirmed and then violated. Going beyond other recent investigations (Li et al., 2011), the authors identify subregion-specific amygdala blood oxygen level-dependent responses that separately reflect outcome prediction and prediction error signals.

Pavlovian fear conditioning, in which initially meaningless conditioned stimuli (CSs) paired with noxious unconditioned stimuli (USs) acquire the ability to elicit fear, has served as a primary model for studying the neurobiological basis of learning. Much of the research generated by that model has been based on variants of the dictum of Hebb (1949), often paraphrased as ‘systems of cells that fire together, wire together’. The amygdala quickly emerged as a site at which CS and US information converged, and hence could be ‘wired together’ when CSs and USs occurred contiguously in time.

However, CS–US contiguity alone is insufficient for associative learning to occur. For example, if a US is already well predicted on the basis of one CS, pairings of a compound of that CS and a new CS with the US often result in little evidence for learning about the new CS, a phenomenon known as ‘blocking’. To deal with many such observations, most learning theories of the past 40 years incorporate the idea that new learning depends critically on prediction error, the difference between expected and received outcomes. Within these models, the importance of CS–US contiguity in the establishment of associations is reaffirmed, but processing of either the CS, the US, or both, is modulated by prediction error, such that unexpected USs or the CSs that precede them (or both) are processed better than expected USs or their accompanying CSs. Considerable evidence from reward conditioning procedures supports the view that the processing of both CSs and USs is indeed modulated by prediction error, and has indicated a number of brain substrates for this modulation, including midbrain dopamine neurons and the amygdala (Holland & Maddux, 2010).

In this study, participants were exposed to a discrimination reversal procedure, in which initially one CS was paired with shock and another CS was not, and later the roles of the two stimuli were reversed. Although a ‘US processing’ model, in which prediction error modulates US effectiveness, fit participants’ ratings of shock expectancy better than a random model, a ‘hybrid’ model that included effects of prediction error on both CS and US processing fared best. The authors examined the amygdala and midbrain blood oxygen level-dependent correlates of CS-induced predictions of shock and its absence, and of the prediction errors induced by the surprising omission or presentation of the shock, when the discrimination was reversed, fitted to the hybrid model. Correlates of unsigned prediction error when the US was unexpectedly presented or omitted were observed in both centromedial amygdala and substantia nigra/ventral tegmental areas, whereas the basolateral amygdala blood oxygen level-dependent response during the CSs was negatively correlated with subsequent prediction error, and hence was related to prediction accuracy.

The work nicely demonstrates convergence of human and animal research concerning fundamental issues of learning in the questions posed (what are the consequences of the confirmation and violation of learned expectancies for information processing), the approaches taken (quantitative modeling based on well-documented theories of learning), and the behavioral and neural processing results obtained, despite differences in species, behavioral measures, and measures of brain activity. The use of common approaches and theoretical perspectives across human and animal studies, each with their own strengths and shortcomings, may provide a unified approach to understanding relations between cognitive and affective processing.