Expectation violation dynamically engages a bottom-up encoding state 1 2

9 Expectation violation has been shown to engage adaptive memory formation, 10 resulting in better memory for unexpected information. In two experiments we tested 11 whether this mechanism is engaged dynamically in a goal-irrelevant manner during 12 retrieval, and how it affects trial-by-trial recognition. Participants encoded images of 13 objects, and then learned a contingency between a cue and category (man-made or 14 natural) with new objects. Targets and parametrically manipulated similar foils, 15 comprising set events, were used at retrieval. In each retrieval trial a cue appeared, 16 which either matched or mismatched (according to the established contingency) the 17 following object, for which participants made an old/new decision. We found that 18 unexpected events at retrieval were associated with increased activation along the 19 ventral visual stream, whereas expected events engaged parietal regions of the core 20 recollection network. For targets and most similar foils, we found an interaction 21 between current and previous expectation status on memory performance, such that 22 expected events following unexpected ones (UprevEcurr) showed a boost in 23 performance. This behavioural effect was associated with activation in the 24 hippocampus, SN/VTA and occipital cortex. A combination of two unexpected events 25 (UprevUcurr) resulted in the poorest memory performance and was associated with 26 increased activation in occipital cortex. Taken together, our findings suggest 27 expectation violation engages an encoding mechanism, supported by bottom-up 28 processing, in a task-independent manner. Therefore, when the goal is to retrieve 29 information, the mnemonic consequences of the shift towards an encoding state is 30 detrimental in real-time, but beneficial for subsequent similar events. 31 32 Abstract word count: 245 33 34


3 8
We found more CR2 for unexpected F2, following F1 and F3 events, compared to  participant was excluded from all analyses due to failure to learn the cue-outcome 2 9 7 contingency during the rule-learning task.

9 8
Procedure. A similar paradigm and expectation manipulation was used in 2 9 9 Experiment 2, with the following exceptions; in the encoding phase, each object was focusing on the details. Following the third presentation there was another jittered 3 0 7 fixation cross, for a longer period of time (800-1200ms), to create mini-blocks 3 0 8 separating each object.

0 9
The rule-learning task was identical to that used in Experiment 1, except a longer objects were not used, as they did not yield any effects of interest in the Experiment 3 1 6 1. Each retrieval trial started with a jittered fixation cross (250-750ms), followed by a 3 1 7 presentation of the cue for 1000ms and then the set event (target, F1 or F2) for 3 1 8 3000ms. In all scanned tasks, we used implicit baselines (fixation crosses for 3 1 9 3500ms in encoding and rule-learning tasks, 4500ms in retrieval) in 30% of trials.

2 0
Behavioural data analysis. Following the analysis and results from Experiment 1, we 3 2 1 collapsed targets and F1 events and modelled the probability of making a correct  and a distractor task, which was not scanned.

9
fMRI data were pre-processed and analysed using SPM12 ( reported are significant at the uncorrected level p < 0.001 with at least 9 contiguous 3 6 4 voxels. Given our a priori hypothesis for the ROIs introduced above (bilateral 3 6 5 hippocampus and SN/VTA), a small volume correction (SVC) approach was 3 6 6 adopted. When SVC was applied, activations were also corrected for family-wise 3 6 7 error (FWE) for the ROI volume.

6 8
To test the behavioural effect of an interaction between a previous event's contextual encoding mini-block (three repetitions of the same item). These estimates were used 3 9 0 as parametric modulators applied to retrieval data in a GLM analysis splitting 3 9 1 retrieval trials by expectation status. Next, we converted the beta estimates of 3 9 2 encoding activity within this mask into t-maps and used them as parametric 3 9 3 modulators applied to retrieval data (using the beta estimates did not change the 3 9 4 results). This approach allowed us to query retrieval data looking for regions whose 3 9 5 retrieval activity correlated with hippocampal encoding activity for expected vs.

9 7
Reinstatement analysis. To identify which regions showed top-down retrieval-3 9 8 oriented activity, we conducted an encoding-retrieval similarity (ERS) analysis using

5 4
Comparing current expected vs. unexpected events, following previously unexpected

6 6
Main effects of expectation. To explore whether expected and unexpected events, 4 6 7 across event types, responses, and positions elicited differential activations in a 4 6 8 bottom-up (ventral visual stream) or reinstatement (recollection network) manner, we 4 6 9 also compared the two conditions using an F contrast (see Figure 4a). We found   observed for first or third set events. c) Schematic of reinstatement and encoding analyses.

9 7
Encoded-related hippocampal activity and encoding-retrieval similarity were used as parametric 4 9 8 modulators (in separate analyses) to query retrieval activation for expected and unexpected events.  In two experiments, we used a contextual expectation manipulation during retrieval 5 2 8 to examine how expectation violation dynamically modulates information processing 5 2 9 and subsequent hippocampal-dependent memory. We found unexpected events mechanism it is necessary to also demonstrate mnemonic consequences of the 5 5 7 increased weight on sensory inputs. Indeed, we observed an interaction between 5 5 8 current and previous expectation status of similar events (targets and F1), such that 5 5 9 memory performance was modulated by a previous unexpected event (improved 5 6 0 recognition performance for U prev E curr , reduced for U prev U curr ).

6 1 6 2
Whilst occipital involvement was observed for both U prev E curr and U prev U curr events 5 6 3 (i.e. irrespective of memory performance), U prev E curr events were also associated with 5 6 4 hippocampal and SN/VTA activation. This finding, together with the memory boost 5 6 5 for U prev E curr events, pertains to the role these regions play in mediating processing 5 6 6 of perceptual inputs for memory formation. In terms of shifting the weight towards Montaldi, 2015). Taken together, these findings suggest that expectation-driven 5 7 3 increased weight on bottom-up inputs is task-independent, but its mnemonic 5 7 4 consequences appear to depend on the task at hand. During learning or exploration,

7 5
further encoding supports later memory for the unexpected event (Garrido, Barnes,