high performance computing on graphics processing units: hgpu.org

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 dissertation investigates the benefits of such parallel models in terms of efficiency and accuracy. First, the viability of a parallel implementation of bio-inspired metaheuristics for function optimization on consumer-level graphic cards is studied in detail. Then, in an effort to expose those parallel methods to the research community, the metaheuristic implementations were abstracted and grouped in an open source parameter/function optimization library libCudaOptimize. The library was verified against a well known benchmark for mathematical function minimization, and showed significant gains in both execution time and minimization accuracy. Crossing more into the application side, a parallel model of the human neocortex was developed. This model is able to detect, classify, and predict patterns in time-series data in an unsupervised way. Finally, libCudaOptimize was used to find the best parameters for this neocortex model, adapting it to gesture recognition within publicly available datasets.