1
Mnemonic prediction errors promote detailed memories
Oded Bein
1*
, Natalie A. Plotkin
2
, Lila Davachi
2,3
Affiliations
1
Princeton Neuroscience Institute, Princeton University
Princeton, NJ, 08540,
2
Department of Psychology, Columbia University
New York, NY, 10027, United States
3
Center for Clinical Research,
The Nathan S. Kline Institute for Psychiatric Research,
Orangeburg, NY, 10962, United States
*Correspondence: oded.bein@princeton.edu
Key words
Prediction error, memory, distinctiveness, pattern separation, mnemonic discrimination,
learning
Running title: prediction errors promote detailed memories
31 pages, 2 figures, 1 Supplementary Information.
Word count, abstract: 254
Word count, main text (excluding abstract, Methods and References): 7,703
2
Abstract
When our experience violates our predictions, it is adaptive to update our knowledge to promote
a more accurate representation of the world and facilitate future predictions. Theoretical models
propose that these mnemonic prediction errors should be encoded into a distinct memory trace to
prevent interference with previous, conflicting memories. We investigated this proposal by
repeatedly exposing participants to pairs of sequentially presented objects (A->B), thus evoking
expectations. Then, we violated participants’ expectations by replacing the second object in the
pairs with a novel object (A->C). The following item memory test required participants to
discriminate between identical old items and similar lures, thus testing detailed and distinctive
item memory representations. In two experiments, mnemonic prediction errors enhanced item
memory: participants correctly identified more old items as old when those items violated
expectations during learning, compared to items that did not violate expectations. This memory
enhancement for C items was only observed when participants later showed intact memory for
the related A->B pairs, suggesting that strong predictions are required to facilitate memory for
violations. Following up on this, a third experiment reduced prediction strength prior to violation
and subsequently eliminated the memory advantage of violations. Interestingly, mnemonic
prediction errors did not increase gist-based mistakes of identifying old items as similar lures or
identifying similar lures as ‘old’. Enhanced item memory in the absence of gist-based mistakes
suggests that violations enhanced memory for items’ details, which could be mediated via
distinct memory traces. Together, these results advance our knowledge of how mnemonic
prediction errors promote memory formation.
3
Introduction
Most of our daily experiences are highly repetitive and predictable. We typically take the same
route to work every day, or we enjoy our favorite soup at the neighborhood restaurant over and
over again. Through repetition, we develop predictions and expectations of what will happen
within a specific context. Once in a while, however, we encounter surprising events that violate
these expectations. For example, we might enter our neighborhood restaurant expecting to have
our favorite soup, but we find out on that evening that the restaurant is offering brussels sprouts
as an appetizer, instead. We term such surprising events “mnemonic prediction errors”: situations
where we expect one thing based on our memory, but the reality is different.
When we encounter a mnemonic prediction error, it is adaptive to update our memories
in order to make better and more accurate predictions in the future. How does memory updating
happen? Interestingly, despite the proposed beneficial role that novelty and prediction errors play
in learning and memory (Ergo et al., 2020; Frank and Kafkas, 2021; Friston, 2018; Henson &
Gagnepain, 2010; Niv & Schoenbaum, 2008; Quent et al., 2021; Reichardt et al., 2020; Rescorla
& Wagner, 1972; Schomaker & Meeter, 2015; Schultz et al., 1997), very little is known about
how mnemonic prediction errors modulate memory encoding. Theoretical models propose that
mnemonic prediction errors should be encoded as distinct memory traces (Frank et al., 2020;
Gershman et al., 2014; Love et al., 2004; McClelland et al., 1995). That is, events that violate our
expectations should be allocated a unique memory representation distinct from prior memories.
This may facilitate memory for the unexpected event, while also mitigating interference with
existing memories that may still be relevant. Indeed, the complementary learning systems
framework shows computationally that the absence of a separated memory trace for mnemonic
prediction errors results in catastrophic interference – incorrectly erasing previous memories
(Kumaran et al., 2016; McClelland et al., 1995). In the restaurant example, this would mean that
immediate integration of the evening that the restaurant served the brussels sprouts instead of
your favorite soup could lead to updating your memory to hold that the restaurant no longer
serves the soup. This might be maladaptive, as the restaurant may still serve the soup on other
nights. Thus, theoretically, memory enhancement of mnemonic prediction errors via a distinct
and separated memory trace enables remembering the event that violated our expectations, while
also protecting previous, potentially relevant memories.
4
However, empirical evidence that mnemonic prediction errors facilitate encoding of
distinct memory traces is scarce. Previous studies provide some evidence that novel and
potentially unexpected events enhance memory (Tulving & Kroll, 1995; von Restorff, 1933).
Most notable is the “von Restorff effect”, or, more broadly, oddball manipulations, whereby a
rarely occurring item that is clearly distinct from the rest of the items in a list (e.g., a dog in a list
of fruits) is better remembered later on (Hunt, 1995; Ranganath & Rainer, 2003; Schomaker &
Meeter, 2015; von Restorff, 1933). One potential account for this effect is that such isolated
items are remembered better because they are unexpected and thus salient (e.g., Axmacher et al.,
2010; Green, 1956; Murty et al., 2016). However, alternative accounts that do not involve
violation of expectations posit that isolation per se can facilitate memory, perhaps because
isolated items elicit less interference during retrieval (Hunt, 2006; Waddill & McDaniel, 1998).
These accounts rely on repeated observations that enhanced memory for isolated items is also
obtained when the isolated item is the second item in the list, before any expectations can be
formed (e.g., Hunt, 1995; Schmidt & Schmidt, 2017). Hence, it is likely that enhanced memory
for oddballs need not result from violation of prior expectations.
Recent studies have directly manipulated violation of expectations and found enhanced
memory for such violations (Antony et al., 2020; Brod et al., 2018; Greve et al., 2017; Kafkas &
Montaldi, 2018a). In one study, Greve et al. (2017) taught participants through repeated exposure
that different scene categories predict either positive or negative valence words, and then
violated this expectation by altering the valence of the words. Memory of the word-scene
association was higher for words that violated prior contingencies (Greve et al., 2017). Similarly,
Kafkas and Montaldi (2018a) taught participants that a specific symbol-cue is followed by either
a man-made object or natural object, and later switched the contingency in some of the trials to
violate participants’ predictions. Recollection rates were higher for objects that violated
participants’ prior expectations, suggesting that mnemonic prediction errors did indeed enhance
memory for unexpected items (Kafkas & Montaldi, 2018a).
Nonetheless, this prior work cannot speak to whether encountering a mnemonic
prediction error results in a detailed memory representation that is potentially distinct from other
memories. This is because enhanced memory can result from either having a distinct
representation of a specific event, or from integrating across memories (e.g., Clewett et al., 2019;
DuBrow & Davachi, 2017; Favila et al., 2016; LaRocque et al., 2013; Schlichting & Preston,
5
2015). For example, it has been shown that items that were later remembered were less similar to
each other (more distinct) in their hippocampal multivoxel activity patterns, compared to items
that were later forgotten (LaRocque et al., 2013), suggesting that distinct item representations
promote memory (see also Favila et al., 2016; Jenkins & Ranganath, 2016). Other studies,
however, have shown that integration across experiences can benefit memory (e.g., Richter et al.,
2016) and that similarity between multivoxel representations of items mediated accurate
temporal memory (DuBrow & Davachi, 2014; see also Schlichting et al., 2015). Thus, it is
currently unknown whether mnemonic prediction errors promote detailed and distinct memories.
In the current behavioral study, we addressed this question by violating participants’
predictions and then conducting a memory test that gauges memory distinctiveness (Bakker et
al., 2008; Stark et al., 2019). To evoke expectations, we employed a statistical learning paradigm
in which participants were repeatedly presented with a stream of objects. Unbeknownst to the
participants, we embedded neighboring pairs of objects within the stream that always appeared in
the same order (Figure 1A). Past work has established that through repetition and learning,
participants come to predict the second object in the pair upon seeing the first object (G. Kim et
al., 2014, 2017; Kok et al., 2017; Schapiro et al., 2012). Following the prediction learning phase,
we violated expectations in half of the pairs by replacing the second object in the pair with a
novel object (Figure 1B). The other half of the pairs were presented intact, thus, no violations
occurred. We scattered novel objects after these intact pairs to serve as a no-violation baseline
condition where no sequential predictions were violated.
Critically, the violation phase was followed by a memory test that targets memory
distinctiveness. We presented the participants with either old items from the violation phase,
similar lures (a different exemplar of an object presented during the violation phase, e.g., a
different pair of scissors than the scissors presented in the violation phase, Figure 1C), or novel
foils that were only presented during the memory test (Figure 1C; Bakker et al., 2008; Lacy et
al., 2011; Stark et al., 2019). The participants indicated whether an item was identical to an item
they had seen before (‘old’), ‘similar’, or ’new’. It is thought that such a fine-grained memory
discrimination requires retrieval of perceptual details of the learned items, to know whether the
item presented during the memory test is identical or, rather, only similar, to the item seen during
learning. Successful discrimination therefore potentially indicates a distinct memory of the
learned items, and can manifest in two ways. The first way is specifically endorsing an old item
