Not what U expect: Effects of Prediction Errors on Episodic Memory.

Abstract

The characterization of the relationship between predictions and one-shot episodic encoding poses an important challenge for memory research. On the one hand, events that are compatible with our previous knowledge are thought to be remembered better than incompatible ones. On the other hand, unexpected situations, by virtue of their surprise, are known to cause enhanced learning. Several theoretical accounts try to solve this apparent paradox by conceptualizing prediction error (PE) as a continuum ranging from low PE (for expectation matching events) to high PE (for expectation mismatching ones). Under such framework, the relationship between PE and memory encoding would be described by a U-shape function with higher memory performance for extreme levels of PE and lower memory for middle levels of PE. In this study we used a gradual manipulation of the strength of association between scenes and objects to render different levels of PE and then tested for episodic memory of the (mis)matching events. In two experiments, and in contrast to what was anticipated, recognition memory as a function of PE followed an inverted U-shape, with higher performance for intermediate levels of PE. Furthermore, in two additional experiments we showed the relevance of explicit predictions at encoding to reveal such inverted U pattern, thus providing the boundary conditions of the effect. We discuss our current findings in the light of the uncertainty in the environment and the importance of the operations underlying encoding tasks.

Accurate predictions facilitate robust memory encoding independently from stimulus probability

“…In schema research, participants are often shown object in a congruent or incongruent context (Quent et al, 2022;Van Kesteren et al, 2013), where incongruent pairs are thought to generate prediction errors. Although some studies have shown better memory for schema-incongruent pairs (e.g., Quent et al, 2022), many have found the opposite result (e.g., Höltje & Mecklinger, 2022;Ortiz-Tudela et al, 2021;Van Kesteren et al, 2013). In addition, it is worth noting that both reinforcement learning and schema research tend to use recognition or forced-choice memory tests (e.g., Greve et al, 2017Greve et al, , 2019, and research has shown better memory for unexpected items during recognition, but not recall (Lew & Howe, 2017).…”

From generating to violating predictions: The effects of prediction error on episodic memory

“…Their results followed the U-shape function of PE, suggesting better memory for predicted and unpredicted events when compared to the medium level. On the other hand, the authors also pointed out that it remains unknown to investigate the U-shape function when predictions are driven from an episodic context rather than pre-experimental knowledge (Ortiz-Tudela et al, 2021).…”