Structuring Memory Through Inference‐Based Event Segmentation

Abstract

Although the stream of information we encounter is continuous, our experiences tend to be discretized into meaningful clusters, altering how we represent our past. Event segmentation theory proposes that clustering ongoing experience in this way is adaptive in that it promotes efficient online processing as well as later reconstruction of relevant information. A growing literature supports this theory by demonstrating its important behavioral consequences. Yet the exact mechanisms of segmentation remain elusive. Here, we provide a brief overview of how event segmentation influences ongoing processing, subsequent memory retrieval, and decision making as well as some proposed underlying mechanisms. We then explore how beliefs, or inferences, about what generates our experience may be the foundation of event cognition. In this inference‐based framework, experiences are grouped together according to what is inferred to have generated them. Segmentation then occurs when the inference changes, creating an event boundary. This offers an alternative to dominant theories of event segmentation, allowing boundaries to occur independent of perceptual change and even when transitions are predictable. We describe how this framework can reconcile seemingly contradictory empirical findings (e.g., memory can be biased toward both extreme episodes and the average of episodes). Finally, we discuss open questions regarding how time is incorporated into the inference process.

Tea With Milk? A Hierarchical Generative Framework of Sequential Event Comprehension

“…This exemplifies a fundamental principle of Bayesian inference known as the Bayesian Ockham's razor (MacKay, 2003, Chapter 28): Although we try to explain incoming data as simply as possible, if we encounter new information that is very unlikely given these prior assumptions, we will always infer (retrieve or learn) the hypothesis that assigns the highest likelihood to the data (see also Shin & DuBrow, 2021, this issue for a discussion in relation to classical rational models of categorization, cf. Anderson, 1991 and Sanborn, Griffiths, & Navarro, 2010).…”

“…In all these cases, we need to be able to infer what kinds of prior experiences are most relevant for the present situation (see Kleinschmidt & Jaeger, 2015, for a detailed discussion of relevant issues). As discussed by Shin and DuBrow (2021, this issue), the Dirichlet process infinite mixture models described above provide insights into how we might be able to learn and extract abstract features that are common to different event clusters, allowing for this type of generalization. These types of models can also explain specific patterns of memory and decision biases (see Franklin et al, 2020).…”

How Does the Mind Render Streaming Experience as Events?

“…As we alluded to earlier, evidence for information optimization in event perception points to a necessary revision to EST. We note that Shin and Dubrow (2021) also suggest important, but different, revisions to EST.…”

“…On this view, event perception is a form of data compression, which can have great psychological value if—as recent findings clearly confirm—the chunks derived from this compression process are amenable to other cognitive operations, such as memory encoding, memory retrieval, hierarchical organization, linguistic labeling, and causal inference (e.g., Brunec, Moscovitch, & Barense, 2018; Buchsbaum et al, 2015; Chekaf, Cowan, & Mathy, 2016; Christiansen, 2019; Christiansen & Chater, 2016; Miller, 1956; Shin & Dubrow, 2021).…”

Events, Event Prediction, and Predictive Processing

“…It may be that inferring an event begins with selecting a discrete model or policy, which minimizes, or is expected to minimize, prediction error, and which can be (passively or actively) tested against sensory input (for discussion, see Baldwin and Kosie, 2021; Shin and DuBrow, 2021). The “coherence” of the interrelated causes of an event would then relate to the actual or expected (local) prediction error minimum achieved by inferring the event—events are then causal models that are particularly good at explaining away actual or expected sensory input.…”