Connectionist Reasoning and Knowledge Representation



For the past several years, my research was mainly concerned with everyday commonsense reasoning by agents. This type of reasoning was characterized by a mixture of rule-based and similarity-based processes, exhibiting both rigor and flexibility (as demonstrated in my AIJ paper). To capture such reasoning, I developed a hybrid connectionist architecture (named CONSYDERR) with both localist and distributed components, that unified rule-based and similarity-based processes and accounted for a variety of CSR patterns.

Within the framework, the following issues were also investigated: (1) The connectionist implementations of rules, logics, and schemas, and the variable binding problem in such implementations. They formed the basis for complex reasoning in connectionist models. (2) Inheritance reasoning, which is an integral part of many CSR patterns. Within CONSYDERR, an intensional approach was developed that works in constant time. This work suggests that other similar reasoning patterns may also be handled intensionally. (3) Causality, which is an important commonsense construct. A connectionist account was developed based on CONSYDERR, which extended the existing logic-based account and dealt better with the inexact, cumulative, and subjective nature of commonsense causal reasoning.

Some attempts have also been made to extend the framework to deal with metaphor and analogy. Further work will be done to refine the architecture and to account for human CSR quantitatively.

Go to Prof. Ron Sun's Homepage