Neuromorphic, or brain-like, electronic systems that mimic cognitive functions are the focus of research because of their potential for complex tasks such as pattern-recognition. Papers presented at the International Electron Devices Meeting in 2011 described studies using programmable phase-change memory (PCM) synapses in neuromorphic systems to carry out a function called spike-timing-dependent plasticity (STDP). STDP is an electronic analog of a brain mechanism for learning and memory, so an electronic system that accurately performs STDP can be said to be “learning.”

At this year’s IEDM, a team led by Korea’s Gwangju Institute of Science and Technology will detail a high-speed pattern-recognition system comprising CMOS “neurons” and an array of resistive-RAM (RRAM)-based “synapses,” which demonstrated STDP. The 1-Kb RRAM array has a simple crosspoint structure and possibly can be scaled to 4F (the theoretical minimum size for a crosspoint array). The work shows the feasibility of using neuromorphic architecture for high-speed pattern recognition.

Photo of CMOS neuron circuit with 1kB RRAM array as synapse

Schematic structure of the proposed system. Input spikes come from the left into the RRAM array. (The inset shows the user interface of a computer simulator.)  The ten input images in the neuromorphic system are learned by edge weighting, and during the learning process ‘5’ in node 4 is represented clearly.

In this comparison of artificial brain projects, Gwangju Institute of Science and Technology’s neuromorphic device is compared to other reported devices.