Hardware Self-Organizing Map With Dithered Step Neighborhood Function

In this paper, a hardware-based SOM architecture with a dithered step (DS) neighborhood function is introduced. In the DS function, dithering is added to a conventional step neighborhood function, and the dithering endows the step function with nonlinear characteristics. The nonlinear characteristics given by the dithering are found to lead to better quantization performance, while the simple structure of the step function, which remains the basis of the proposed function, maintains a low hardware cost. Here, the characteristics of the dithered function were theoretically analyzed before clustering experiments utilizing the proposed SOM implemented on a field-programmable gate array (FPGA) were conducted to evaluate the extent to which the proposed neighborhood function improves the performance of the resulting SOM. The experimental results show that the proposed DS neighborhood function decreases the circuit size and power consumption by 13.86% and 17.61%, respectively. In addition, the DS neighborhood function was used to create a hardware-based SOM with 4.33% greater throughput.