Bio-inspired approach for information fusion.
2017-02-14T00:16:55Z (GMT) by
Sensory cue integration in the human brain plays a crucial role in perception. The neurobiology community have assembled an ample amount of data and behavioral evidence in recent decades enhancing our understanding about sensory information processing in the human brain and in perception. The primary aim of this thesis is to model the sensory cue integration in the biological brain by drawing on the neurobiological findings gathered so far. At the higher level of abstraction, four main characteristics of sensory cue integration are examined this research endeavor by means of modeling; multimodal dynamics, hierarchical sensory cue integration, crossmodal interactions and multisensory integration. Moreover, the computational modeling effort addresses the limitations of some current fusion techniques and theoretical frameworks on multisensory integration. Four unsupervised neural network models are designed and implemented to capture the characteristics of sensory cue integration. Eventually a fusion framework is built by merging the functionalities and neuronal dynamics of these four neural network models. Empirical evaluation of the these models are performed using real data sources drawn from the application ~eas such as image processing, lip reading and bioinformatics. Further analysis of these models explains some of the key neurobiological properties in cross¬modal integration: binding problem, crossmodal matching, congruent/incongruent cue processing, reliability-based selective attention, inverse effectiveness, silent lip reading and ensemble coding hypothesis in cue integration. While these models explain possible mechanisms underlying sensory integration they can be used to solve problems in real data mining applications. The fusion framework allows proper integration of multiple data sources to discover hidden patterns and rela¬tionships in an unsupervised manner. Furthermore, it allows a data analyst to investigate patterns under different level of granularities from multiple data sources.