Improved UFHLSNN (IUFHLSNN) for Generalized Representation of Knowledge and Its CPU Parallel Implementation Using OpenMP

Abstract

Fuzzy Hyper Line Segment Neural Network (FHLSNN) (Kullarni et al., International Joint conference on 4:2918–2933, 2001) is a hybrid system that combines fuzzy logic (Zadeeh, IEEE Trans. Fuzzy Syst. 4:103, 1996) and neural networks (Zurada, Fundamental Concepts and Models of Artificial Neural Systems, 1992, pp. 30–36). It is used extensively for real-world pattern classification (Zurada, Fundamental Concepts and Models of Artificial Neural Systems, 1992, pp. 30–36). It learns patterns in terms of n-dimensional Hyper Line Segment (HLS). Modified Fuzzy Hyper Line Segment Neural Network (MFHLSNN) (Patil et al., The 12 IEEE International Conference, vol. 2, 2003) is a modified version of FHLSNN (Kullarni et al., International Joint conference on 4:2918–2933, 2001) that improves the quality of reasoning and recall time per pattern using modified fuzzy membership function. Updated Fuzzy Hyper Line Segment Neural Network (UFHLSNN) (Dhabe, 2016 International Conference on Computing, Analytics and Security Trends, 2016) for larger pattern datasets is proposed using minimum computational efforts to compute membership. In this chapter, we proposed improved version of UFHLSNN (Dhabe, 2016 International Conference on Computing, Analytics and Security Trends, 2016), called IUFHLSNN, for generalized representation of knowledge for better recognition. IUFHLSNN uses midpoints of HLSs computed for the recall phase and thus expected to provide better recognition, as suggested in Occam’s razor principle (Blumer et al., Inform. Process. Lett. 24:377–380, 1987).

We compared serial and parallel implementations using Intel’s Xeon E5-2620 and obtained average speedup of 16.96× and 77.22×, respectively, for classification and recognition, for all the used datasets (Poker Dataset, https://archive.ics.uci.edu/ml/datasets/Poker+Hand; QtyT40I10D100K DataSet, https://archive.ics.uci.edu/ml/datasets/QtyT40I10D100K; Skin Segmentation Dataset, https://archive.ics.uci.edu/ml/datasets/skin+segmentation). In the same experiment the obtained percentage gain in time are 92% and 94%, respectively, for classification and recognition. Thus, we strongly recommend IUFLSNN and its OpenMP parallel execution. We also compared parallel executions on two different computing CPU architectures, viz. IBM’s POWER8 and Intel’s Xeon E5-2620. We found that IBM’s POWER8 is two times faster than Intel’s Xeon E5-2620. Thus, we strongly recommend generalized representation of HLS knowledge and OpenMP parallelization.