Contribution
A review that states event segmentation theory (EST) in information processing terms, maps its components onto brain systems, and gathers behavioural, narrative, neuropsychological and imaging evidence. It joins the older unitization literature (Newtson’s button-pressing task) and narrative situation-model work to a single mechanism: event boundaries are moments of transient prediction error at which the current event model is reset.
Key insight
“Perceptual systems continuously make predictions about what will happen next. When transient errors in predictions arise, an event boundary is perceived” (Abstract). An event model is “a representation of “what is happening now,” which is robust to transient variability in the sensory input” (A Theory of Event Segmentation). Its input is gated: closed most of the time, open briefly when error rises. “Thus, the system alternates between long periods of stability and brief periods of change.”
Assumptions
- Perception is hierarchical, recurrent and cyclical, and oriented to predicting the near future (the “working assumptions” of many theories).
- Event models are working memory. “Event models are working memory representations, which are implemented by transient changes in neural activation rather than long-term changes in synaptic weights.” Event schemata, by contrast, are long-term knowledge that biases event models continuously.
- One global reset signal per timescale, plausibly dopamine or norepinephrine, rather than selective updating of parts of a model; this is what distinguishes EST from gating models of prefrontal cortex.
- The perceiver is passive. The authors list this as a limitation: “A second limitation is that the theory in its current form is a purely passive perceiver.”
Key results
A theory and review; no new data except a cited network simulation.
- The mechanism. Updating “consists of (1) resetting the current representations … and (2) transiently increasing the influence of the pathway from sensory inputs to the event models”. It runs on several timescales at once, coarse resets roughly a subset of fine ones.
- Two control roles. Segmentation allocates processing over time (“event segmentation does not itself require attention; rather, it implements a mechanism of attention”) and updates working memory by resetting event models. Hence its claim to be “a core, domain-general mechanism of cognitive control”.
- Six implications listed: segmentation is spontaneous and needs no conscious attention; it is cognitive control; it is multi-timescale; multimodal; it “depends on change. When the world is static, prediction is easy”; and it depends on prior knowledge.
- A partial simulation (Reynolds, Zacks & Braver, then in press): a network predicting body positions of an actor shows larger prediction error at event boundaries, and error-gated event-model units improve prediction without labels. Event schemata are not implemented.
- Behavioural evidence. Unitization is reliable within and across observers; coarse boundaries align with fine ones; boundaries are better remembered; segmenting well predicts later memory, in older adults with and without Alzheimer’s dementia; edits and commercial breaks that respect boundaries help recall and those that cut across them hurt it; movement change and goal completion predict boundaries.
- Narrative evidence. Situation-model dimensions (time, space, objects, characters, causes, goals) predict where readers mark boundaries, more so as more dimensions change. In Speer & Zacks (2005) a time shift (“an hour later”) slowed reading and lowered recognition of what came just before. This is the paper’s direct evidence for its memory prediction: “Thus, information presented prior to an event boundary should be difficult to retrieve following that boundary.”
- Neural mapping. Transient responses at boundaries in extrastriate cortex (near MT+) and right posterior frontal cortex during passive viewing; anterior cingulate cortex proposed for prediction error; midbrain neuromodulatory nuclei for the reset; lateral prefrontal cortex for event models and schemata (lesion, Parkinson’s, schizophrenia and script-imaging studies).
- The admitted misfit. “The theory makes one prediction that does not square well with the available data”: increased ACC activity at event boundaries during passive viewing has not been observed. Prefrontal cortex does not generally respond at boundaries during passive viewing either.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | People segment ongoing activity into events reliably, on several timescales at once, with coarse boundaries aligned to fine ones | strong as cited | unitization studies (Newtson; Zacks, Tversky & Iyer 2001; Hard et al.) |
| C2 | Event boundaries are perceived at transient rises in perceptual prediction error | moderate: a simulation and correlational evidence; prediction error not measured in people | Reynolds et al. network; predictability ratings of clauses at changes |
| C3 | At a boundary, event models (working memory) are reset, so information from before it becomes harder to retrieve | moderate | Speer & Zacks 2005 (time shifts in narrative); related reading-time studies |
| C4 | Boundaries receive extra processing and are better remembered; segmentation quality predicts memory | moderate to strong as cited | Newtson & Engquist 1976; Boltz 1992; Schwan & Garsoffky 2004; Zacks et al. on dementia |
| C5 | Segmentation is spontaneous, needs no conscious attention, and is a mechanism of attention and cognitive control | moderate: passive-viewing fMRI responses at later-marked boundaries | Zacks et al. 2001; Speer et al. 2003 |
| C6 | Event models and schemata are implemented in lateral PFC; prediction error in ACC; reset by midbrain neuromodulators | weak: a tentative mapping, partly contradicted | lesion and imaging studies; the ACC misfit |
Concepts
- Event: “a segment of time at a given location that is perceived by an observer to have a beginning and an end” (after Zacks & Tversky 2001); the paper’s core cases are goal-directed human activity lasting seconds to tens of minutes.
- Event model: working-memory representation of what is happening now, stable against transient input, gated open at boundaries. The authors say it is “analogous to the term “situation model” used in the narrative comprehension literature”.
- Event schema: long-term, learned knowledge of the sequential structure of kinds of activity; continuously biases event models.
- Unitization: the button-press task for marking boundaries, kept distinct from segmentation as an ongoing process.
- Gating: the mechanism that opens event models to sensory input when prediction error rises and closes it as error falls.
Connections
It descends from Zacks & Tversky (2001), Newtson’s unitization work, and narrative situation-model theory (Zwaan & Radvansky 1998; the event indexing model). It compares itself to adaptive resonance theory and to prefrontal gating models (Miller & Cohen 2001; Frank, Loughry & O’Reilly 2001), differing in that error is perceptual, not reward error. It does not mention doorways or spatial navigation tasks; the doorway studies (LIT-1392 and after) build on the same situation-model literature, and the later ones cite EST’s follow-up papers. The record’s active-inference account of conscious content (THEORY-883) shares the prediction-error architecture; EST makes no claim about consciousness, and says event models “are not necessarily accessible to consciousness”.
Bearing on the record
- CASE-136. The case calls EST “the theory behind” the doorway effect. This paper does not discuss the doorway effect; the record’s identification of A37’s “Event Segmentation Theory” with this paper is the record’s own, and it is a fair one. But EST and the account the doorway authors give (LIT-1384, LIT-1385) are not the same mechanism. EST predicts that what came before a boundary becomes harder to retrieve because the event model is reset. The doorway effect falls on the object being carried into the new event, which a reset of the old model should spare. That is why the doorway authors turn to interference between models. The case lists both as if one theory explained the effect.
- CLAIM-421 (activation interruption) and A37’s reading. A37 read the doorway effect through EST: the intention “may have lost accessibility because the current situation is being organized through a different event model”. As a statement of EST that is fair: a reset event model no longer holds what it held. But EST is a theory of perceiving others’ activity. Event models in it represent the world, not one’s goals, and the authors name extending it to “how event models explicitly represent one’s own goals” as future work. An intention dropped at a boundary is an extension of EST, not a consequence of it.
- TERM-055. EST’s split between event models (working memory, reset at boundaries) and event schemata (long-term, continuous) is a mechanism for the difference between holding and having to hand: knowledge persists in schemata while the model that made it available is reset. The paper does not put it that way.
- THEORY-688 (doorway options in learning). The options framework segments action at subgoals; EST segments perception at prediction error. EST’s own comparison is with prefrontal gating models, which gate on reward error; options are not discussed.
- THEORY-886 is filed from this reading (Proposed): the core of EST. The record held no account of event segmentation.
- No instruction for machine-learning practice.
Limitations
- Prediction error is never measured in people. The behavioural evidence links boundaries to change, goals and statistical structure, all of which could drive boundaries without error as the intermediary. The one direct test of the error mechanism is a simulation on simple motion-capture actions.
- Its one neural prediction checked (ACC at boundaries) failed, and the neural mapping is offered as tentative.
- Much of the cited evidence is from the authors’ groups, and several key studies were in press or under review at the time.
- The memory-cost prediction rests chiefly on narrative time shifts, not on experienced events.
- It is a theory of a passive observer of goal-directed human activity.
Open questions
- Would manipulating predictability while holding perceptual change fixed move boundaries, as C2 requires? (Later work by the same group addresses this; not read here.)
- Does EST predict the doorway effect’s specific pattern, a cost on the carried object? Or does that need a second mechanism, such as the event horizon model’s interference?
- How does an actor’s own goal enter an event model, so that an intention could be lost at a boundary?