Parallel Bio-Inspired Methods for Model Optimization and Pattern Recognition
Nature based computational models are usually inherently parallel. The collaborative intelligence in those models emerges from the simultaneous instruction processing by simple independent units (neurons, ants, swarm members, etc...). This talk will present the benefits of such parallel models in terms of efficiency and accuracy. First, the viability of a parallel implementation of bio-inspired meta-heuristics for function optimization on consumer level graphic cards is described in detail. Then, in an effort to expose those parallel models to the research community, the meta-heuristic implementations were abstracted and grouped in an open source parameter/function optimization library, libCudaOptimize. This library was proven effective in various applications from different fields. Varying from detecting pathological structures in medical images to finding traffic signs in videos obtained from autonomous vehicles. It has also shown significant gains in both execution time and optimization accuracy, thanks to its parallel implementation.
The talk will also present a novel parallel model of the human neocortex. This model, termed HQSOM, is able to detect, classify, and predict patterns in time-series data, in an unsupervised way. Experimental results for a gesture recognition task show promising results on videos acquired by the Microsoft Kinect sensor. Moreover, the model was tested successfully on different modalities of data, ranging from audio signals, to handwritten digits classification problems.