Statistical prediction of the future impairs episodic encoding of the present

Abstract

Memory enables reminiscence about past experiences and guides processing of future experiences. However, these two functions are inherently at odds: remembering specific past experiences requires storing idiosyncratic properties, but by definition such properties may not be shared with similar situations in the future and thus are not as useful for prediction. We discovered that, when faced with this conflict, the brain prioritizes prediction over encoding. Behavioral tasks showed that pictures allowing for prediction of what will appear next based on learned regularities are poorly encoded into memory. Brain imaging revealed that predictive representations in the hippocampus may be responsible for this worse episodic encoding and suggested that such interference may result from competition between hippocampal pathways. This tradeoff between statistical learning and episodic memory may be adaptive, focusing encoding on experiences for which we do not yet have a predictive model. whereas statistical learning refers to extracting what is common across mul-5 tiple experiences (e.g., what tends to happen at birthday parties). Episodic memory allows for vivid recollection and nostalgia about past events, whereas statistical learning leads to more generalized knowledge and predictions about new situations. Episodic memory occurs rapidly and stores even related expe-10 riences distinctly in order to minimize interference, whereas statistical learning occurs more slowly and overlays memories in order to represent their common elements or regularities. Given these behavioral and computational differences, theories of memory have argued that these two kinds of information must be processed serially and stored separately in the brain (McClelland et al., 1995; 15 Squire, 2004): episodic memories are formed first in the hippocampus and these memories in turn provide the input for later statistical learning in the neocortex as a result of consolidation (Frankland and Bontempi, 2005; Richards et al., 2014;Tompary and Davachi, 2017).Here we reveal a relationship between episodic memory and statistical learn-20 ing in the reverse direction, whereby learned regularities determine which memories are formed in the first place. Specifically, we examine whether the ability to predict what will appear next -a signature of statistical learning -reduces encoding of the current experience into episodic memory. This hypothesis depends on two theoretical commitments: first, that the adaptive function of memory is 25 to guide future behavior by generating expectations based on prior experience ; second, that memory resources are limited, because of attentional bottlenecks that constrain encoding (Aly and Turk-Browne, 2017) and/or because new encoding interferes with the storage or retrieval of existing memories (Shiffrin and Atkinson, 1969). Accordingly, in allocating memory re-30 sources, we propose that it is less important to encode an ongoing experience when it already generates strong expectations about future states of the world...

Evidence of hippocampal learning in human infants

“…In particular, the first item of each learned pair in structured blocks may trigger a hippocampal process of prediction or pattern completion of the second item. 34,35 In addition to clarifying the timing and consequences of hippocampal participation in infant statistical learning, future research is needed to determine whether this role is necessary for behavioral expression of statistical learning. This will be difficult to test in infants, but studies in adult patients with hippocampal damage suggest that the hippocampus may in fact be necessary for normal statistical learning behavior.…”

When to retrieve and encode episodic memories: a neural network model of hippocampal-cortical interaction

“…Relatedly, a recent paper by Bein et al (2020) found that prediction error (which is closely related to surprise) biased hippocampal dynamics towards encoding, and other work has found that prediction error boosts memory (e.g., Greve et al, 2017; see also Henson and Gagnepain, 2010). In addition to surprise (and prediction error), recent work by Sherman and Turk-Browne (2020) suggests that predictive certainty may play a role in shaping encoding policy: They found that stimuli that trigger strong predictions (i.e., high certainty about upcoming events) are encoded less well. In keeping with this point, Bonasia et al (2018) found that, during episodic encoding, events that were more typical (and thus were associated with more predictive certainty, and less surprise) were associated with lower levels of medial temporal lobe (MTL) activation.…”

Prediction Errors Disrupt Hippocampal Representations and Update Episodic Memories

“…The sample size was determined a priori to satisfy the following conditions: (1) reproduce the sample size and achieve at least 90% power to detect the interaction effect found in prior work (ηp 2 = 0.17) (39), and (2) evenly 11 allocate participants to 6 pseudorandomized trial order lists. Participants were healthy young 12 adults (age: M = 22.42, SD = 2.41, range [18,30]; gender: 75% female, 25% male). Inclusion 13 criteria were as follows: between the ages of 18-30, normal or corrected-to-normal vision and hearing, no history of neurological or psychiatric disorders, and fluency in English.…”

“…What mechanisms could link hippocampal prediction errors to memory updating? A leading hypothesis is that prediction errors shift the focus of attention and adjust cognitive 22 processing (20,(28)(29)(30)(31)(32). After episodes that align with expectations, we should continue 23 generating predictions and shift attention internally, sustaining and reinforcing existing 1 memories.…”