A Neural Network Approach to Model-Based Recognition
Abstract
The problem of recognizing structured patterns (object recognition) is being pursued on several fronts. The main thrust of the work involves the design of networks that store some notion of a relational model of an object and performs recognition via a version of graph matching. This approach is governed by the use of objective functions to both specify the network and the problem to be solved. The dynamics of the net thus carry out an optimization procedure. Key here is the incorporation into the objective function of compositional and specialization hierarchy of models, and provision to perform dynamic grouping (perceptual organization) of the input data. Results so far show very good performance for versions where data is preprocessed into a form matchable to the database, but poorer performance on more difficult problems where the network must itself organize raw data into relational structures for matching. A related effort explores aspects of traditional associative memories that may be of use in more complex networks. Questions of performance, storage and robustness are addressed. A new fast learning algorithms is proposed for a CMAC network. Work in optical implementation of some of these networks constitutes a third front. The main problem here is to use optics to form a fixed interconnection network between layers of 2-D nodes (neurons). Two means of using spatial multiplexing to effect a 4-D interconnect between two 2-D node planes are used: multifaceted holograms and multichannel incoherent image systems.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 1989
- Accession Number
- ADA206990
Entities
People
- Gene R. Gindi
Organizations
- Yale University