Construction and Training of Multi-Associative Graph Networks

Adrian Horzyk, Daniel Bulanda, Janusz A. Starzyk

Abstract: Modern methods and networks of supervised learning use a vast amount of computational resources when adapting to large datasets. They are unable to incorporate new training examples into trained models quickly and to represent internal knowledge for quick adaptation to other computational tasks without retraining. The human brain is capable of representing and retrieving vast amounts of information and can create associations between its various pieces based on frequent relationships and patterns. The backpropagation algorithm is not the best and only way to train neural networks, especially since its use is limited to feed-forward architectures. Brain structures can be modeled using graph architectures that are not feed-forward but recursive with many feedback connections. This paper introduces Multi-Associative Graph Networks that enable the representation of associated training data and objects transformed from relational databases. These graphs store the data along with the most useful relationships to facilitate computational intelligence processes. We describe the associative transformation algorithm allowing for the transformation of any relational database into this graph network, reproducing stored relationships and enriching them with newly detected ones. We also introduce the tuning algorithm that learns to associate different priorities with different neurons representing objects to improve the relational dependencies and classification results. Finally, we draw conclusions from the comparisons to other state-of-the-art models.