Dimensionality Reduction Techniques to Aid Parallelization of Machine Learning Algorithms

In order to handle the complexity of issues encountered in wireless networking and signal processing, recently researchers have resorted to employing machine learning techniques. This paper proposes the use of dimensionality reduction techniques to aid in the parallelization of machine learning algorithms. For an example use case, we first examine a mode of data parallelism that is based on Radon's theorem. We then bring out the bottleneck of this scheme in training models on high dimensional datasets and demonstrate the advantage of incorporating dimensionality reduction techniques like principal component analysis (PCA) to improve its performance. We provide theoretical justification for the improvement in the performance. The proposed algorithm was implemented on an NVIDIA GTX 1060 graphics processing unit using PYCUDA python libraries. We evaluate the performance of the proposed algorithm using skin segmentation, HEPMASS, HIGGS and SUSY datasets and bring out the trade off between f-scores and execution time.