Design Space Exploration for K-Nearest Neighbors Classification Using Stochastic Computing

K-nearest neighbors (KNN) algorithm is a powerful classification approach in various machine learning and pattern recognition tasks. As a non-parametric algorithm, KNN is viewed as one of the most efficient machine learning algorithms and hence it is widely adopted in various practical applications. However, due to its extensive use of high-complexity multiplication, to date the efficient hardware architecture design of KNN is still very limited, thereby impeding the potential applications of KNN in real-time embedded systems. This paper explores the design space of efficient low-complexity KNN. By utilizing the emerging stochastic computing (SC) technique, the hardware cost of KNN can be significantly reduced. In particular, multi-line-based SC representation schemes and low-cost dimension reduction approach are proposed to increase the test accuracy of KNN with high variance source data. Analysis results show that the proposed KNN designs achieves good balance between high classification accuracy and low hardware cost.