Temporal - spatial recognizer for multi-label data

• Main Author: Mousa, Aseel
• Format: Thesis
• Language: English; English; English
• Published: 2018
• Subjects: TK7885-7895 Computer engineering. Computer hardware
• Online Access: https://etd.uum.edu.my/7439/1/Depositpermission_s93596.pdf; https://etd.uum.edu.my/7439/2/s93596_01.pdf; https://etd.uum.edu.my/7439/3/s93596_02.pdf; https://etd.uum.edu.my/7439/; http://sierra.uum.edu.my/record=b1697792~S1
• Institution: Universiti Utara Malaysia

Pattern recognition is an important artificial intelligence task with practical applications in many fields such as medical and species distribution. Such application involves overlapping data points which are demonstrated in the multi- label dataset. Hence, there is a need for a recognition algorithm that can separate the overlapping data points in order to recognize the correct pattern. Existing recognition methods suffer from sensitivity to noise and overlapping points as they could not recognize a pattern when there is a shift in the position of the data points. Furthermore, the methods do not implicate temporal information in the process of recognition, which leads to low quality of data clustering. In this study, an improved pattern recognition method based on Hierarchical Temporal Memory (HTM) is proposed to solve the overlapping in data points of multi- label dataset. The imHTM (Improved HTM) method includes improvement in two of its components; feature extraction and data clustering. The first improvement is realized as TS-Layer Neocognitron algorithm which solves the shift in position problem in feature extraction phase. On the other hand, the data clustering step, has two improvements, TFCM and cFCM (TFCM with limit- Chebyshev distance metric) that allows the overlapped data points which occur in patterns to be separated correctly into the relevant clusters by temporal clustering. Experiments on five datasets were conducted to compare the proposed method (imHTM) against statistical, template and structural pattern recognition methods. The results showed that the percentage of success in recognition accuracy is 99% as compared with the template matching method (Featured-Based Approach, Area-Based Approach), statistical method (Principal Component Analysis, Linear Discriminant Analysis, Support Vector Machines and Neural Network) and structural method (original HTM). The findings indicate that the improved HTM can give an optimum pattern recognition accuracy, especially the ones in multi- label dataset.