Novelty, Prediction Error and Memory Encoding: Limitations of the Pimms Framework

Change log
Quent, Jörn Alexander 

The Predictive Interactive Multiple Memory Systems (PIMMS) framework has been used to explain how novelty, or more precisely “prediction error”, boosts memory encoding. In this thesis, I explored several other phenomena in the animal and human literature that PIMMS cannot yet explain but should. PIMMS predicts that unexpected information will be better encoded than expected information. However recent work has suggested that expected information can also be better remembered than less expected information. By using a range of expectancies for the location of objects with an immersive virtual reality (iVR) kitchen, I showed that memory is a “U-shaped” function of expectancy, with best memory for highly expected or highly unexpected locations relative to intermediate levels of expectancy. Using OSF-registered Bayesian inference, this U-shape was consistent across four experiments. While the advantage for highly unexpected locations is consistent with PIMMS, the advantage for highly expected locations is not. Importantly, the advantage for expected locations was not simply due to a guessing bias when the location was forgotten, suggesting that the advantage arises during encoding rather than just at retrieval. This U-shape is consistent with another framework - the SLIMM framework - which proposes that different brain regions support the two ends of the U-shape, such that the advantage for unexpected information should be associated with recollection of contextual information via a medial temporal lobe system (like in PIMMS), while the advantage for expected information should be associated with a feeling of familiarity based on rapid cortical consolidation enabled by a medial prefrontal cortex system. However, when I asked participants to indicate recollection or familiarity at retrieval, both ends of the U-shape continuum were associated with higher recollection, while there was no detectable effect of expectancy on familiarity. I consider why this SLIMM prediction may therefore be incorrect. Another finding in the literature concerns the effect of novelty on unrelated information shortly preceding or succeeding the novel experience. PIMMS says nothing about this penumbra effect, which has been related to plasticity-related proteins triggered by the novel experience (so-called “behavioural tagging”). Since participants report that their first iVR experience is highly novel, I submitted a Registered Report to test whether iVR affected memory for unrelated words that were encountered prior to entering the iVR room. In short, the finding was that there is no evidence that novelty improves memory performance for information learned before experiencing something novel. Possible reasons for the failure of finding an effect were discussed. A final limitation of PIMMS I considered was the effect of “boundaries” in continuous stimuli, which are known to affect memory for the temporal order of information. While boundaries might be generated by prediction errors, PIMMS is silent on how they affect temporal order memory. Using a movie featuring a series of rooms, I tested whether memory for the temporal order of objects encountered in those rooms is affected by doorways between rooms and/or by surprising/perceptual changes within a room. Unfortunately, I was unable to replicate a previous report where temporal order memory was worse for pairs of objects in different rooms (i.e., either side of a doorway) than objects in the same room, let alone either sides of a surprising/perceptual change within a room. Taken together, my findings and the literature demonstrate the multiple potential factors that determine how novelty affects memory encoding (and consolidation), which require a more comprehensive theoretical framework than currently available.

Henson, Richard
memory, prediction error, surprise, novelty
Doctor of Philosophy (PhD)
Awarding Institution
University of Cambridge