Recurrent correlation associative memories

• Main Author: Teo, Mei Ping
• Other Authors: Quek Hiok Chai
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/59993
• Institution: Nanyang Technological University

The online technique of neuro-fuzzy system has been increasing in popularity in the recent years. In actuality external factors play an important role in the time-variant dataset, changing its pattern. This change in pattern is known as drift and shift. To tackle these changes, Hebbian learning was introduced. However this learning is characterised by uni-directional learning, resulting in the instability of the model. Hence, the BCM theory was developed to overcome the problem of Hebbian learning through the provision of Hebbian and Anti-Hebbian learning. However, time variant data possesses both dynamic and temporal problems. The purpose of the author is to address this issue through the modification of the current recurrent fuzzy neural network. The underlying principle is to store past information to be recalled later for application in the current context. The existing recurrent neuro-fuzzy system shows promising results that motivates the author to further the efficacy of the recurrent neuro-fuzzy system. This report proposes a recurrent neuro-fuzzy system that uses the BCM theory of online learning with self-organizing effectiveness. In addition, rules are represented using the Takagi Sugeno Kang model to achieve a better accuracy compared to the Mamdani model which focuses on interpretability. The performance of Recurrent SeroTSK is evaluated and compared against neuro-fuzzy systems through various time-series benchmark experiments and prediction for cancer diagnosis which is a classification data. The results show that Recurrent SeroTSK is better for time-series prediction and it works for classification data as well.