Exploiting Training Example Parallelism with a Batch Variant of the ART2 Classification Algorithm
Ciechanowicz P, Dlugosz S, Kuchen H, Müller-Funk U
In this article we develop a batch variant of the ART 2 clas siﬁcation algorithm invented by Carpenter and Grossberg. Our algorithm exploits training example parallelism while leaving the overall design of the ART 2 network unchanged such that a signiﬁcant reduction of the execution time can be achieved on a multiprocessor system. We present a par allel implementation strategy and analyze it w.r.t. execu tion time and speedup. As our algorithm naturally beneﬁts from data parallelism, the implementation uses data paral lel skeletons of the Muenster skeleton library Muesli. We show that skeletons are an efﬁcient way to write parallel applications compared to a manual MPI implementation.
ART 2, batch, skeletons, training example parallelism