Learning in Structured Connectionist Networks
Abstract
Connectionist networks compute in a manner analogous to real neural networks. The work in this thesis focuses on computing and especially learning in structured connectionist networks -- those which emphasize problem-specific connection patterns as well as adaptive weight change rules. A connectionist chart parser was implemented which parses limited-length strings for context- free grammars in a constant number of parallel computations steps. The parser was extended to disambiguate and complete near-miss parses, as well as learn new production in certain limited situations. While the parser works well, the structure is too rigid and learning too difficult for cognitive modeling. Two algorithms for learning simple, feature-based concept descriptions were also implemented. The first recruits hidden units representing pairs of features. The performance of this network is good when the definition involve pairs of input features, but attempts to build hierarchies of pair units for longer definitions were not successful. The second algorithm is an enhancement of an existing technique, competitive learning, adding feedback from concept units to guide the partitioning of inputs into classes. This technique is more successful and is used as a component in a network which is capable of learning descriptions of structured objects.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1988
- Accession Number
- ADA206852
Entities
People
- Mark A. Fanty
Organizations
- University of Rochester