Bio inspired self-organizing map based architecture for improved capture of data sequence

The advancement of ICTs has enabled higher prevalence of sequential data generated by various fields of study such as engineering, biology and finances. The availability of such sequential data has generated significant interest in both the research community and businesses to develop models and techniques to capture,analyze and learn from sequential data. Although models generated from sequential information could provide a plethora of useful information that are not visible in non-sequential data, learning sequences remains challenging task. The handling of variable length associated with sequences, identifying substructure of a sequences and pattern mining (e.g. motifs) from sequences are the three mains challenges in the contemporary context. Nevertheless, we humans have the ability to comprehend and utilize information from complex sequences such as speech and vision in an effortless manner. In fact, the ability of the human brain to store and utilize sequences represent a key element of human intelligence. Such an ability is directly attributed to the information propagation within the human neocortex and the concept storage in neurons via stronger synaptic bonds. By considering these two factors as an inspiration, this thesis initially formulates a conceptual model for sequence learning and identifies seven properties that are essential for a successful implementation of such a model. By utilizing these properties as an evaluation metric, SOM was chosen among others as the most suited base algorithm for the development of the sequence capturing model. The SOM was extended to process sequential information by enabling the temporal processing of information via modeling of inputs as matrices that preserves the temporal ordering of the each data event. Subsequently, the extended SOM was utilized as the base map on which additional layers were introduced to the SOM map by exploiting the winner stabilization property SOM learning. The layered model is initially implemented to learn discrete sequences by employing a single additional layer to the SOM map followed by learning of continuous sequences via discretization using additional layers. Each layer represents an abstraction of information represented by the layer below it and the knowledge was modeled as state machines in the newly introduced layers. The number of states and the transitions incident on a state machine decreases with each additional layer since the states from immediate bottom layer is merged at the next higher layer via grouping techniques to enable a greater understanding of the dataset. To evaluate the implemented model, six datasets consisting of three mitochondrial DNA (mtDNA) datasets, two geometric datasets and a purpose collected game-play dataset were used. The mtDNA datasets were used to generate the reference phylogenies from the subsequences extracted from the model using the average common substring (ACS) like method. The game-play dataset was used to identify preferred path sequences of players when playing the predator-prey game Pac-Man whilst the geometric datasets was used to evaluate the models ability to learn continuous sequences. Finally, a running time analysis was performed to evaluate the processing time of the developed model by comparing it with the ACS method and the recently introduced ASTrie and ASTree algorithms.