Cognitive connectionist models for recognition of structured patterns

• Main Author: Wong, James Jia Jun
• Other Authors: David Cho Siu-Yeung
• Format: Theses and Dissertations
• Language: English
• Published: 2008
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Pattern recognition
• Online Access: https://hdl.handle.net/10356/14263
• Institution: Nanyang Technological University

Traditional pattern recognition by computers focuses on the problem of identifying simple two-dimensional templates, such theories are too simplistic to account for the human‟s abilities to recognize varied and novel patterns. Feature theories ignore evidence that processing of global form often takes priority over processing of local features and are sensitive to context in which the stimulus appears. Pattern recognition systems usually consist of three steps of data acquisition, feature extraction and classification. Feature extraction process in pattern recognition, produces errors, more than often, is due to the operating environment that the feature extractor is used. Typically, a recognized object can be subjected to various degrees of changes. This motivates us to develop another kind of feature representations for pattern recognition. Many natural or artificial systems are more appropriately modelled using “Data Structures”. By incorporating structures in extracted features, it would facilitate the data processing process and later pattern recognition process by making it more efficient and noise tolerant. This thesis is presented to investigate the use of connectionist models to generalize structural information, which perform like a human cognition for recognizing erratic patterns. Erratic patterns here mean that incomplete features are extracted by feature extractor in a pattern recognition system, caused by occlusions in the data or un-filterable noise in the pattern. A computational framework for learning a flavour of structural connectionist models is of paramount importance for both pattern recognition and development of brain-inspired systems, since it allows the treatment of structured information very naturally and, in several cases, very efficiently. The details of this framework will be investigated in this thesis.