LAUSR

In recent years Reservoir Computing has arisen as an emerging machine-learning technique that is highly suitable for time-series processing. Nevertheless, due to the high cost in terms of hardware resources, the implementation of these systems in one single chip is complex. In this brief, we propose a hardware implementation of a reservoir computing system with morphological neurons that allows us to reduce considerably the area cost associated with the neural synapses. The main consequence of using tropical algebra is that input multipliers are substituted by adders, leading to much lower hardware requirements. The proposed design is synthesized on a Field-Programmable Gate Array (FPGA) and evaluated for two classical time-series prediction benchmarks. The current approach achieves significant improvements in terms of energy efficiency and hardware resources, as well as an appreciably higher precision compared to classical reservoir systems.