SoC Kohonen Maps Based on Stochastic Computing

Mobile systems and by extension Internet of Things (IoT) applications requests more and more Machine Learning functions, thus requiring a big computational power with a small power available. These demands have led to renewed interest in unconventional hardware computing methods capable of to implement complex functions in a simple way, with a very small power consumption. This work proposes a novel System-on-Chip (SoC) implementation of a Kohonen Map based on stochastic computing. In turn, to support this development, several stochastic block designs are presented as Winner-Take-All (WTA) similarity check and the squared Euclidian distance. The capabilities and performance of the proposed SoC solution is tested over a well-known classification task over the Fisher's Iris data set, archiving the same classification performance than the software. The proposed solution requests few hardware resources and low power, due to its inherent capacity to implement complex functions in a simple way. This enables to implement large self-learning classifiers based on Kohonen maps on tiny systems.