Nearest neighbor classification on two types of SIMD machines

Abstract

The nearest neighbor (NN) classification schemes have been popular in pattern recognition because of their efficiency and simplicity. The computation of distances of a testing pattern to the training patterns is very time-consuming. This paper proposes several parallel algorithms for nearest neighbor classification on two different types of SIMD machines. One type of SIMD machine has parallel shared memory modules. An alignment network is used to connect the processors and the memory modules. The data storage schemes and two different procedures on this machine type are presented. The other type of machines employs only a distributed memory system and an interconnection network. The data distribution and communication issues are discussed. For a problem with N p training patterns and N f features, comparing the distances from a testing pattern to all training patterns takes O( N pN f ) time on a sequential machine. The distance computation and comparison can be performed in parallel. The proposed algorithms on a parallel machine with Q processing elements (PEs) take approximately O( N pN f/Q ) time.