A neural net model for discrete mappings foundations and implementation

Thesis/Dissertation · Sun Jan 01 04:00:00 EST 1989

OSTI ID:6779034

Koutsougeras, C

Neural net models suggest an interesting approach towards development of adaptive machines. A new neural net model is presented which can automatically adapt its input-output behavior to any mapping from a discrete vector space onto a finite set of elements, the mapping being completely or incompletely specified in terms of an example set of mapping pairs. The proposed model introduces a new type of unit (neuron) and a standard tree-like feed forward interconnection topology. The network is intended to automatically build internal representations of classification rules or concepts which are adequate for discriminating between two classes of input vectors. The descriptions of the classification concepts are internally expressed by separating hypersurfaces. Internal representations are developed during a training phase. Expansion of the network takes place during training so as to guarantee a sufficient clustering of the input space. Both the network's expansion and the decisions, made locally at each unit during the training phase, are guided by an entropy measure which is appropriately defined. Unlike models in which the network topology is specified before training, in this model the network expands during training. This model therefore, does not face the severe problem of deciding what network size is adequate for adaptation to a certain desired input-output behavior, a problem which plagues existing models. The outlines of three architecture versions are also presented which realize the model. An implementation of the model by means of a clustered connectionist architecture is described. A second implementation is also described which is based on data flow concepts. The second architecture is shown to accommodate novel advances in optical processing in an excellent way. A third architecture version employing optical processing structures is also presented.

• Research Organization: Case Inst. of Tech., Cleveland, OH (USA)
• OSTI ID: 6779034
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
990200* -- Mathematics & Computers
990300 -- Information Handling
AUTOMATION
COMPUTER ARCHITECTURE
DATA-FLOW PROCESSING
DECISION TREE ANALYSIS
IMPLEMENTATION
INFORMATION SYSTEMS
MAPPING
MATHEMATICS
NEURAL NETWORKS
OPTICAL MODES
OSCILLATION MODES
PROGRAMMING
TOPOLOGICAL MAPPING
TOPOLOGY
TRANSFORMATIONS
VECTOR PROCESSING