Figure 2.

Temporal Context Model (TCM): overview of operation and representations. (A) The TCM: Howard et al., 2005: there are two main layers (Item, Context). Training trials are presented as input to the item layer. M^FT indicates connections between item and context layer. M^TF indicates connections between context and item layer. Hebbian learning typically used to update these connections. See main text for details. (B) Similarity (i.e., overlap) of representations over the context layer induced by each object captures direct (e.g., A-B) and indirect associations (e.g., A—C). White shading indicates a value of zero, and black a value of one. Prior to the experiment, the contextual states activated by each objects are assumed to be orthogonal to each other—denoted by black shading only across the diagonal (upper figure). After the study phase, the contextual representations induced by objects in a triplet (A, B, C) have now become similar to each other—but not to objects in other triplets (e.g., X, Y, Z). Critically, this is the case for contextual representations associated with A and C objects [i.e., denoted by shading in [A, C] square: even though these objects were presented in different study pairs (i.e., AB, BC)]. It is worth noting that the similarity between objects presented in the same study pair (e.g., A, B) is greater than that between different pairs (i.e., A, C). Note that the similarity between different objects reflects the dot product between vectors specifying the contextual representation associated with each object. See Howard et al., 2005 for implementational details of the TCM model.