Training and operation of an integrated neuromorphic network based on metal-oxide memristors

• Published in: Nature (London) Vol. 521; no. 7550; pp. 61 - 64
• Main Authors: Prezioso, M., Merrikh-Bayat, F., Hoskins, B. D., Adam, G. C., Likharev, K. K., Strukov, D. B.
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 07.05.2015; Nature Publishing Group
• Subjects: 631/378/3920; 639/166/987; 639/925/927/1007; Algorithms; Biomimetics; Cerebral cortex; Electronics - instrumentation; Engineering; Equipment Design; Humanities and Social Sciences; Humans; Integrated circuits; letter; Memristors; Metals - chemistry; Models, Neurological; multidisciplinary; Nanotechnology; Neural Networks (Computer); Oxides - chemistry; Properties; Science; Semiconductor chips; Semiconductors; Synapses - physiology; Transistors, Electronic
• ISSN: 0028-0836; 1476-4687
• DOI: 10.1038/nature14441

A transistor-free metal-oxide memristor crossbar with low device variability is realised and trained to perform a simple classification task, opening the way to integrated neuromorphic networks of a complexity comparable to that of the human brain, with high operational speed and manageable power dissipation. A neuromorphic network based on metal-oxide memristors Building neuromorphic networks matching the cognitive complexity of their biological prototypes but with higher performance is one of the great challenges in computing. One promising approach to such devices — potentially simpler than those based on complex silicon circuits — combines complementary metal-oxide-semiconductors (CMOSs) with adjustable two-terminal resistive devices (memristors). Here Dmitri Strukov and colleagues demonstrate a transistor-free metal-oxide memristor network with low device variability that works as a single-layer perceptron. That is, it can learn to recognize imperfect 3 × 3 pixel black-and-white patterns as one of three letters of the alphabet. The strength of this approach is its scalability so that larger neuromorphic networks capable of more challenging tasks should be possible. Despite much progress in semiconductor integrated circuit technology, the extreme complexity of the human cerebral cortex 1 , with its approximately 10 14 synapses, makes the hardware implementation of neuromorphic networks with a comparable number of devices exceptionally challenging. To provide comparable complexity while operating much faster and with manageable power dissipation, networks 2 based on circuits 3 , 4 combining complementary metal-oxide-semiconductors (CMOSs) and adjustable two-terminal resistive devices (memristors) have been developed. In such circuits, the usual CMOS stack is augmented with one 3 or several 4 crossbar layers, with memristors at each crosspoint. There have recently been notable improvements in the fabrication of such memristive crossbars and their integration with CMOS circuits 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , including first demonstrations 5 , 6 , 12 of their vertical integration. Separately, discrete memristors have been used as artificial synapses in neuromorphic networks 13 , 14 , 15 , 16 , 17 , 18 . Very recently, such experiments have been extended 19 to crossbar arrays of phase-change memristive devices. The adjustment of such devices, however, requires an additional transistor at each crosspoint, and hence these devices are much harder to scale than metal-oxide memristors 11 , 20 , 21 , whose nonlinear current–voltage curves enable transistor-free operation. Here we report the experimental implementation of transistor-free metal-oxide memristor crossbars, with device variability sufficiently low to allow operation of integrated neural networks, in a simple network: a single-layer perceptron (an algorithm for linear classification). The network can be taught in situ using a coarse-grain variety of the delta rule algorithm 22 to perform the perfect classification of 3 × 3-pixel black/white images into three classes (representing letters). This demonstration is an important step towards much larger and more complex memristive neuromorphic networks.