Memristor-Based Neuromorphic System with Content Addressable Memory Structure
By mimicking the complex biological systems, neuromorphic system is more efficient and less energy-efficient than the traditional Von Neumann architecture. Due to the similarity between memristor and biological synapse, many research efforts have been inv
- PDF / 596,559 Bytes
- 10 Pages / 439.37 x 666.142 pts Page_size
- 109 Downloads / 241 Views
2
School of Automation, Huazhong University of Science and Technology, Wuhan 430074, China [email protected] Key Laboratory of Image Processing and Intelligent Control of Education Ministry of China, Wuhan 430074, China 3 Texas A&M University at Qatar, Doha, Qatar
Abstract. By mimicking the complex biological systems, neuromorphic system is more efficient and less energy-efficient than the traditional Von Neumann architecture. Due to the similarity between memristor and biological synapse, many research efforts have been investigated in utilizing the latest discovered memristor as synapse. This paper improves the original network circuit based on memristor and content addressable memory structure and extends the existing results in the literature. The competition network circuit includes input layer, synapse and output layer. The synapse is made up of two memristors which store information and judge whether input and storage data are same. The output layer consists of subtractor which processes match and mismatch voltage to recognize pattern and the winner-take-all circuit to find out of which storage pattern is the closest to input pattern. The circuit design about read/write framework and working principle are discussed in detail. Finally, the system has been trained and recognizes these 5 × 6 pixel digit images from 0 to 9 successfully. Keywords: Memristor recognition
1
·
Neural network
·
Neuromorphic
·
Pattern
Introduction
In recent years, the concept of memristor, originally proposed by Leon Chua in 1971 [1], has generated renewed interest since it has been experimentally found in 2008 by HP Lab [2]. Memristors and memristive devices are basically resistor with varying resistances, which depend on the history of the current through the memristor. Moreover, memristor-based memories can achieve high integration density of 100 Gb/cm2 [3]. Because of these advantages, they have been thought as a potential candidate for synapses on neuromorphic computing systems and many works have been done in this field. For instance, the memristor was used to represent the connection weight of two neurons utilizing memristor-based synaptic neuron in [4,5]. In [6,7], neural synapse was comprised of transistor c Springer International Publishing Switzerland 2016 L. Cheng et al. (Eds.): ISNN 2016, LNCS 9719, pp. 681–690, 2016. DOI: 10.1007/978-3-319-40663-3 78
682
Y. Zhu et al.
and memristor and the neural network demonstrates fundamental properties including associative learning and pulse coincidence detection. On the other hand, memristor-based neural network with basic learning rule called winner-take-all(WTA) has attract much extensive concern. Ebong et al. designed memristor-based neural network for position detection through the WTA learning rule [8]. In [9], a neuromorphic system which consists of memristor array and CMOS neuron uses same learning rule for visual pattern recognition. In [10], neuromorphic crossbar circuit based on memristors recognize speech through this rule. In studying memristor-based neural ne
Data Loading...
