Shape Adjustable Stochastic Neighborhood Function for Hardware Self-Organizing Map

In order to save implementation costs, hardware self-organizing maps (SOMs) employ simplified neighborhood functions at the expense of reduced performance. This paper proposes a stochastic (STOC) hardware-friendly neighborhood function with shape adjustability without less performance reduction. First, characteristics of the STOC function was theoretically analyzed, and clustering performance of SOM with the STOC function was investigated by simulation. The STOC function was optimized in terms of quantization error by utilizing its shape adjustment feature, and as a result, the STOC-SOM achieved the best quantization performance after the Gaussian function. Then, we evaluate the hardware cost of the STOC function.