Parallel implementations for the Online Sequential Extreme Learning Machine algorithm applied to the Particulate Material

Abstract

The Online Sequential Extreme Learning Machine algorithm is suitable for forecasting Data Streams with Concept Drifts. Nevertheless, forecasting requires high-performance implementations due to the high incoming samples rate. In this work, we analyzed parallel implementations for the Online Sequential Extreme Learning Machine in the C programming language, with OpenBLAS, Intel MKL, and MAGMA libraries. Both OpenBLAS and Intel MKL provide functions that explore the multithread features in multicore CPUs, which expands the parallelism to multiprocessors architectures. In turn, MAGMA offers functions that run in parallel in heterogeneous/hybrid architectures, like Multicore systems with graphics processing unit, the GPU. Thus, the goal of this work is to compare the performance -- prediction error/precision and real stream processing time -- of the C implementations with the original Online Sequential Extreme Learning Machine in MATLAB when forecasting concentrations of Particulate Matter in the air. Experimental results showed that in most cases approached here, at least one of the C implementations obtained better performance, regarding stream processing time, when compared with the reference MATLAB version, performing up to 7-fold faster.