Preface

*We are accused of going against the times. We are doing that deliberately and with all our strength.* —Lanza del Vasto

Neural networks significantly deal with a large area of applications such as image process‐ ing, speech recognition, natural language processing, and bioinformatics. Unfortunately, it is still difficult to fully analyze the inference provided by a layered neural network, as it contains complex parameters embedded in hierarchical layers.

Therefore, nowadays scientific research deals with alternative solutions for analyzing neural network architectures where the stochastic nature and live dynamics of memristive models play a key role. The features of memristors make it possible to direct processing and analy‐ sis of both biosystems and systems driven by artificial intelligence, as well as to develop plausible physical models of spiking neural networks with self-organization.

This book deals with advanced applications illustrating these new concepts, and delivers an important contribution for the achievement of the next generation of intelligent hybrid bio‐ structures.

Different modeling and simulation tools can deliver an alternative to funding the theoretical approach as well as practical implementation of memristive systems.

> **Dr. Calin Ciufudean** Associate Professor "Stefan cel Mare" University Suceava, Romania

**Section 1**

**Memristor Artificial Synapsis for Neural**

**Networks**

**Memristor Artificial Synapsis for Neural Networks**

**Chapter 1**

**Provisional chapter**

**Synaptic Behavior in Metal Oxide-Based Memristors**

With the end of Moore's law in sight, new computing paradigms are needed to fulfill the increasing demands on data and processing potentials. Inspired by the operation of the human brain, from the dimensionality, energy and underlying functionalities, neuromorphic computing systems that are building upon circuit elements to mimic the neurobiological activities are good concepts to meet the challenge. As an important factor in a neuromorphic computer, electronic synapse has been intensively studied. The utilization of transistors, atomic switches and memristors has been proposed to perform synaptic functions. Memristors, with several unique properties, are exceptional candidates for emulating artificial synapses and thus for building artificial neural networks. In this paper, metal oxide-based memristor synapses are reviewed, from materials, properties, mechanisms, to architecture. The synaptic plasticity and learning rules are described. The electrical switching characteristics of a variety of metal oxide-based memristors are

With the aid of modern technology, human society has entered into a new big data era. Meanwhile, it brings a new challenge to humans for data processing. Despite the great success in the past decades, the traditional computer based on Von Neumann architecture and complementary metal oxide semi-conductor (CMOS) technology is still suffering limitations of dealing with big data while it can only deal with well-defined data. These machines cannot compete with the biological system in solving the imprecisely specified problems of the real world which are very simple for biological beings [1, 2]. Even though the digital computers can emulate some functionality

discussed, with a focus on their application as biological synapses.

**Keywords:** memristor, metal oxide, synapse, neuromorphic computing,

**Synaptic Behavior in Metal Oxide-Based Memristors**

© 2016 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

© 2018 The Author(s). Licensee IntechOpen. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use,

distribution, and reproduction in any medium, provided the original work is properly cited.

DOI: 10.5772/intechopen.78408

Ping Hu, Shuxiang Wu and Shuwei Li

Ping Hu, Shuxiang Wu and Shuwei Li

http://dx.doi.org/10.5772/intechopen.78408

**Abstract**

synaptic plasticity

**1. Introduction**

Additional information is available at the end of the chapter

Additional information is available at the end of the chapter

#### **Synaptic Behavior in Metal Oxide-Based Memristors Synaptic Behavior in Metal Oxide-Based Memristors**

DOI: 10.5772/intechopen.78408

Ping Hu, Shuxiang Wu and Shuwei Li Ping Hu, Shuxiang Wu and Shuwei Li

Additional information is available at the end of the chapter Additional information is available at the end of the chapter

http://dx.doi.org/10.5772/intechopen.78408

#### **Abstract**

With the end of Moore's law in sight, new computing paradigms are needed to fulfill the increasing demands on data and processing potentials. Inspired by the operation of the human brain, from the dimensionality, energy and underlying functionalities, neuromorphic computing systems that are building upon circuit elements to mimic the neurobiological activities are good concepts to meet the challenge. As an important factor in a neuromorphic computer, electronic synapse has been intensively studied. The utilization of transistors, atomic switches and memristors has been proposed to perform synaptic functions. Memristors, with several unique properties, are exceptional candidates for emulating artificial synapses and thus for building artificial neural networks. In this paper, metal oxide-based memristor synapses are reviewed, from materials, properties, mechanisms, to architecture. The synaptic plasticity and learning rules are described. The electrical switching characteristics of a variety of metal oxide-based memristors are discussed, with a focus on their application as biological synapses.

**Keywords:** memristor, metal oxide, synapse, neuromorphic computing, synaptic plasticity

#### **1. Introduction**

With the aid of modern technology, human society has entered into a new big data era. Meanwhile, it brings a new challenge to humans for data processing. Despite the great success in the past decades, the traditional computer based on Von Neumann architecture and complementary metal oxide semi-conductor (CMOS) technology is still suffering limitations of dealing with big data while it can only deal with well-defined data. These machines cannot compete with the biological system in solving the imprecisely specified problems of the real world which are very simple for biological beings [1, 2]. Even though the digital computers can emulate some functionality

© 2016 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. © 2018 The Author(s). Licensee IntechOpen. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

of certain animals with comparable speed and complexity, the energy consumptions increase exponentially as the animal hierarchy becomes higher with a very huge volume. Conversely, the biological brain is a compact dense system which can offer parallel processing, self-learning, and adaptivity with a combination of storage and computation in very low power consumption [3]. In these decades, the implementation of Von Neumann architecture computers to mimic biological systems has been in the form of software but such simulations are not comparable to biological systems in terms of efficiency and speed due to the physical limitation of those digital computers. Even the artificial neural networks based on CMOS-integrated circuits are far inadequate for constructing bionic systems. The truly reason for this drawback is the need to transfer data between a memory(storing data) and a processor(computing based on the data). This requirement of data transfer generates an intrinsic delay and inefficiency, which is a bottleneck for all CMOS-based neural networks [4]. In the past decades, the semi-conductive technology has led to great progress under the aid of the rapid development of the electronic industry, which has promoted the steps forward to develop artificial neural networks. In 2011, the supercomputer Watson, with 2880 computing cores [5], won the human-machine contest which proved that supercomputers have their advantages in some aspects [6]. But the important point that has been ignored in this comparison is the energy consumption and the physical volume of the computers. Watson has thousands of cores and requires about 80 kW of power and 20 tons of air-conditioned cooling capacity [7], while the human brain occupies space like a soda bottle and consumes power of 10 W.

**2. Plasticity and learning of the synapse**

neuron connects with other neurons through 10<sup>3</sup>

**Figure 1.** A schematic illustration of synapse.

In the nervous system, neurons and synapses are the basic units for transit information to the whole biological body. In the human brain, it consists of ~1011 neurons and an extremely large number of synapses, ~10<sup>15</sup>, which act as a highly complex interconnection network among the neurons [29]. Neurons consist of three main parts: a soma, dendrites, and an axon. Neurons generate action potentials (spikes), with amplitudes of approximately 100 mV and durations in the range of 0.1–1 ms in their soma. The spikes propagate through the axon and are transmitted to the next neuron through the synapses. A synapse [30] is a 20–40 nm junction between the axon and the dendrites (shown in **Figure 1**) that permits a neuron (or nerve cell) to pass an electrical or chemical signal to another neuron or to the target efferent cell. Each

–10<sup>4</sup>

The information transmission between neurons with the synapses is very complicated which

synapses to form a complex network.

Synaptic Behavior in Metal Oxide-Based Memristors http://dx.doi.org/10.5772/intechopen.78408 5

Therefore, an alternative approach to building a brain-like or neuromorphic computational system with distributed computing and localized storage in networks becomes an attractive option [1, 8–11]. The brain-like computational system can outperform conventional computers with good performance in handing the real-time processing of unstructured sensory data, such as image, video or voice recognition, navigation, etc. [12–17]. Also, the brain-like computational system has the advantages of architecture and function compared to conventional computers, offering massive parallelism, small area, scalability, power efficiency, the combination of memory and computation, self-learning and adaptivity [3]. Many researches have helped us understand how neurons and synapses function and revealed how essential synapses are to biological computations, especially in memorizing and learning [18–21]. However, building compact neuromorphic computing systems remains as a challenge, especially for the lack of electronic elements which could mimic the biological synapses. In recent decades, the research of neuromorphic systems is renewed by the understanding of biological neural networks and the emergence of new nanodevices. Particularly, the emergence of the fourth electronic element, memristor [22–28], makes it feasible to construct bionic hardware which will lead to effective, high-performance neuromorphic computing hardware.

In this chapter, we will discuss synaptic devices and summarize the recent progress in neuromorphic hardware, which is based on memristors. In particular, we will focus on a few typical devices based on metal oxides and their key properties served as synapses. We will start with a brief description of memory and the learning of synapses in Section 2. In Section 3, we will elaborate more on these oxide-based memristors (TiOx , WOx , HfOx, TaOx, NiOx, etc.) with an emphasis on resistive switching (RS) characteristics, which is followed by neuromorphic computing applications and the underlying physical mechanism. We limit this review to metal oxide-based memristive devices for the emulation of synaptic functionalities and will not cover the literature on neuromorphic circuits.
