A FPGA-based Learning Accelerator for Self-Organizing Map and Its Application to Trend-Visualization

Self-Organizing Map (SOM) is one of the neural networks employed, particularly in data mining and related fields. Like Deep Learning, SOM requires iterative calculations of a vast number of high-dimensional vectors. Consequently, dedicated hardware is essential for accelerating SOM computations. We propose a dedicated SOM accelerator utilizing FPGA. In this accelerator, all reference vectors of SOM neurons are implemented in the internal memory (UltraRAM) embedded in the FPGA, and those vectors are calculated in parallel based on the internal parallelism within SOM. Furthermore, we propose a SOM application that visualizes similarities among trends in vast numbers of tweets in X (Twitter), which requires on-line learning that keeps up the rapid changes of tweets. The proposed accelerator is applied to this application, demonstrating approximately 5 times faster performance compared to an Intel Core i7 (multithreaded) processor.