Integrating top-down expectations with bottom-up perceptual processing in a hybrid neural-symbolic architecture

Cognitive actions are frequently taken with the expectation of a particular result in mind. The expectation is generally based on past experiences, but not necessarily on any specific experience in isolation; rather, the integrated collection of past experiences forms the basis for what we should expect. Although an important part of action execution, this aspect of cognition has largely been ignored in many architectures. Here, we demonstrate the implementation of expectation generation by using ACT-R's blending mechanism. Rather than returning a single particular chunk from memory, as ACT-R's normal declarative retrieval does, the blending mechanism returns a chunk whose slot values are generated via a weighted average over all the chunks in the set that matches the request, with the activations of the chunks themselves serving as the weights. In past work, this retrieval mechanism has been used to implement prototype generation from memory as well as to generate certainty measures from metacognitive signals. We embed the expectation generation capability within the larger SAL (Synthesis of ACT-R and Leabra) framework, a hybrid framework developed to take advantage of the relative strengths of the two architectures, and discuss directions for future work, including error monitoring and recovery.