Knowledge acquisition is governed by striatal prediction errors

Discrepancies between expectations and outcomes, or prediction errors, are central to trial-and-error learning based on reward and punishment, and their neurobiological basis is well characterized. It is not known, however, whether the same principles apply to declarative memory systems, such as those supporting semantic learning. Here, we demonstrate with fMRI that the brain parametrically encodes the degree to which new factual information violates expectations based on prior knowledge and beliefs—most prominently in the ventral striatum, and cortical regions supporting declarative memory encoding. These semantic prediction errors determine the extent to which information is incorporated into long-term memory, such that learning is superior when incoming information counters strong incorrect recollections, thereby eliciting large prediction errors. Paradoxically, by the same account, strong accurate recollections are more amenable to being supplanted by misinformation, engendering false memories. These findings highlight a commonality in brain mechanisms and computational rules that govern declarative and nondeclarative learning, traditionally deemed dissociable.


Supplementary Figure 3. Relationship between PE, subsequent accuracy, and subsequent confidence, for incorrect Test1 trials in the recognition study.
A significant relationship was observed between PE on Test1 for incorrectly answered questions and Test2 accuracy, as well as Test2 confidence for the subset of incorrect-tocorrect trials. As PE magnitude increased, efficacy of learning from feedback was superior for both measures (see Results in main text under subheadings 'PE magnitude in Test1 correlates with subsequent memory' and 'Confidence in updated memories correlates with PE'). Displayed are group median values ±SEM.

Supplementary Figure 4. Relationship between confidence updating and Test1 PE for trials answered correctly in both tests of the recognition study.
A. A significant linear relationship was observed between PE on Test1 for correctly answered questions (100-confidence) and confidence expressed in the same (correct) answers in Test2. Notably, the slope of this relationship was significantly less than 1, indicating that as Test1 confidence increased (and positive PE decreased from 90 to 0), the degree of confidence updating decreased. Thus, questions answered correctly with low confidence in Test1 (high positive PE) were associated with large increases in confidence in Test2, whereas for higher confidence correct trials (low positive PE), this metric gradually diminished and even became negative, reflecting a decrease in confidence in Test2 (despite having confirmatory feedback in Test1). This result served as a validation of the PE hypothesis, over an alternative hypothesis that the updating effects observed in the incorrect trials could be explained by positing that low confidence results from greater difficulty of learning that information. Note, there were no Test1 trials answered correctly with zero confidence (equivalent to +100 PE). Displayed is group average confidence in Test2 (±SEM) for each confidence level in Test 1, along with the best fitting linear function. B. The delta, or change in average confidence from Test1 to Test2.

Test1 correct, false feedback trials, in the recall study.
There were three possible Test2 answers for this subset of trials: false feedback Test2 answers were instances where the original correct memory was replaced by the false feedback information; original feedback Test2 answers were instances where the correct Test1 answer was again supplied (i.e. memory not overwritten by false feedback); incorrect Test2 answers were instances where no answer was supplied, or an answer was supplied which did not correspond to the original information, or the false feedback.
Greater PE induced by false feedback (following medium-confidence correct answers) was associated with an increased likelihood of false feedback recollection in Test2 (above), and -in the subset of correct-to-false feedback trials -with greater confidence in those false memories (below). Depicted are group means ± SEM.

Supplementary Figure 6. Two putative models of PE encoding.
FMRI analyses of PE related neural responses were designed to identify brain regions exhibiting BOLD responses (represented as blue lines) consistent with an encoding of the absolute magnitude (salience), or magnitude and valence of the PE in each trial. The two models can be differentiated by the hypothetical BOLD response to negative PEs (in red), which are expected to be positive in the Magnitude model and negative in the Magnitude + Valence model.

Supplementary Table 1. Regions correlating with PE magnitude and valence.
Talairach co-ordinates of the peak voxel, cluster size and t values for the regions showing significant correlative activity with signed PE at a threshold of p < .005 (see Figure 3 and Results in main text under subheading 'Regions correlating with both valence and magnitude of PE'). The right-hand columns indicate which regions were also predictive of subsequent accuracy and confidence in Test2 (see Figure 5 and Results in main text Talairach co-ordinates of the peak voxel, cluster size and t values for the regions showing significant correlative activity with unsigned PE at a threshold of p < .005 (see Figure 4 and Results in main text under subheading 'Regions correlating with absolute PE magnitude'). Trial type as a percentage of the total number of incorrect answers in Test1. A significant number of false feedback answers were supplied in the recall study to determine false memory adoption following correct responses. Since erroneous answers entailed no/incorrect memory, we did not distinguish between false and original feedback when assessing memory updating, i.e. Test2 answers were scored with respect to the feedback supplied in Test1, not the original material. The updating effect for incorrect Test1 trials was also significant when excluding false-feedback trials from the analysis.