Reservoir Computing (RC) is a type of Recurrent Artificial Neural Networks(RANN) that has simple training algorithm. The recurrent connections of the reservoir layer gives the network the ability to process the spatiotemporal signals . Two types of RC are deeply studied at the NanoComputing research Lab, Liquid State Machine(LSM) and Echo State Networks (ESNs). The software implementations of the ESNs has been effective at divers set of applications such as Speech Recognition, weather forecasting and emotion recognition. However, custom hardware implementations are in demand to meet the requirements of some embedded system applications such as therapeutic devices and body sensors, where power consumption, processing speed and area of the system are critical.
A memristive based hardware architecture of Echo State Networks is presented. This architecture uses memristive devices to build current mode synapse circuits. These circuits use two memristive devices to save the weight and has low power consumption compared with other CMOS synapse circuits. A current mode with sigmoid activation function neural circuit is also used. The architecture connects the neuron on a 2-d mesh network by a set of switches and multiplexers which allows reconfiguring the connections of the reservoirs based on different topologies. Two reservoir topologies (Rig and Random) were applied on this architecture.
The architecture was successfully used to design an epileptic seizure detection system with up to 97% accuracy.