A theory of the discovery and predication of relational concepts.

Abstract

Relational thinking plays a central role in human cognition. However, it is not known how children and adults acquire relational concepts and come to represent them in a form that is useful for the purposes of relational thinking (i.e., as structures that can be dynamically bound to arguments). The authors present a theory of how a psychologically and neurally plausible cognitive architecture can discover relational concepts from examples and represent them as explicit structures (predicates) that can take arguments (i.e., predicate them). The theory is instantiated as a computer program called DORA (Discovery Of Relations by Analogy). DORA is used to simulate the discovery of novel properties and relations, as well as a body of empirical phenomena from the domain of relational learning and the development of relational representations in children and adults.

Representing composed meanings through temporal binding

“…This means that synchronous binding cannot be used to accurately represent a concept such as dark pink tree, in which dark directly modifies pink but is screened off from tree (i.e., dark pink serves as a compound adjective). Under a model such as Doumas et al (2008), representing this sort of structure would have to be done using asynchronous binding (e.g., dark and pink would need to be activated at one timestep, followed by tree at a second timestep [ Figure 6]) meaning that the resulting representation would have twice the complexity.…”

“…One possibility is that we flexibly bring each binding scheme to bear depending on the representational demands of the task (Doumas et al, 2008). Asynchronous binding could be used when certain elements must be screened off, e.g., representing dark pink tree as (dark pink) tree.…”

“…However, it displays an important difficulty accounting for certain types of more complex compositional representations. In particular, when using binding by synchrony it is not possible to represent the precise structure with which concepts should be bound (Doumas, Hummel, & Sandhofer, 2008). While the simultaneous activation of a set of concepts does indicate which of them should be bound together, it does not indicate which concepts should serve as arguments and which should serve as predicates.…”

“…An alternative approach, and the one that we focus on here, is that concepts may be screened off from each other by using binding via asynchrony, in which the pattern of activation of concepts over time distinguishes different thoughts and creates constituent structure. For example, a phrase like dark pink tree can be represented with a constituent structure of the form [[dark & pink] tree] by initially co-activating dark and pink (to indicate a dark pink colour), and then subsequently activating tree in isolation (to indicate that the bound concept "dark pink" should modify tree) (Doumas et al, 2008;Hummel & Holyoak, 2003). In this case, the initial period of activation would first activate features associated with darkness and pinkness (which would be well matched by a dark pink colour), and then features associated with treeness (which would be well matched by a tree).…”

“…More complex and structured concepts, however, must be represented by activating the different components of a concept across time. That is to say, the system pays a cost for precisely representing structure (e.g., some models limit the number of timesteps available (Doumas et al, 2008); this provides an upper bound on working memory capacity).…”