Figure 2. A situation-dependent event processing task.

(A) An event is a sequence of states, sampled from an event schema and conditioned on a situation. An event schema is a graph where each node is a state. A situation is a collection of features (e.g., Day, Weather, Sleepiness) set to particular values (e.g., Day = weekday). The features of the current situation deterministically control how the event unfolds (e.g., the value of the Day feature controls which Barista state is observed). At each time point, the network observes the value of a randomly selected feature of the current situation, and responds to a query about what will happen next. Note that the order of queries is fixed but the order in which situation features are observed is random. If a situation feature (Sleepiness) is observed before the state it controls (Drink), the model can answer the query about that state by holding the relevant feature in working memory. However, if the relevant situation feature (Weather) is observed after the state it controls (Mood), the model can not rely on working memory on its own to answer the query. (B) We created three task conditions to simulate the design used by Chen et al., 2016: recent memory (RM), distant memory (DM), and no memory (NM); see text for details. (C) Decoded contents of the model’s working memory for an example trial from the DM condition. Green boxes indicate time points where the value of a particular situation feature was observed. The color of a square indicates whether the correct (i.e., observed) value of that feature can be decoded from the model’s working memory state (white = feature accurately decoded; black = feature not decoded). See text for additional explanation.