What Is an Artificial Neural Network?

Artificial Neural Network (Artificial Neural Network, ANN) is a research hotspot that has emerged in the field of artificial intelligence since the 1980s. It abstracts the human neuron network from the perspective of information processing, establishes some simple model, and forms different networks according to different connection methods. In engineering and academia, it is often referred to as neural network or neural network. A neural network is a computing model that consists of a large number of nodes (or neurons) connected to each other. Each node represents a specific output function, called an activation function. The connection between each two nodes represents a weighted value for the signal passing through the connection, called the weight, which is equivalent to the memory of an artificial neural network. The output of the network varies depending on how the network is connected, the weight value and the excitation function. The network itself is usually an approximation of an algorithm or function in nature, or it may be an expression of a logical strategy.

In the past ten years, the research work of artificial neural network has been continuously deepened, and great progress has been made.It has successfully solved many problems in the fields of pattern recognition, intelligent robots, automatic control, prediction and estimation, biology, medicine, and economics. The practical problems that modern computers are difficult to solve have shown good intelligent characteristics.

TA says

2018-09-13 21:09 "Stupid" neural network2018-09-13 21:09

If life just follows the monotonous rules, then the image recognition function of the neural network must work well. But humans, even sheep, can do unexpected things, and artificial intelligence has become less reliable since then. ... more

Artificial neural network neuron

as the picture shows

a1 ~ an are the components of the input vector

w1 ~ wn is the weight of each synapse of the neuron

b is the offset

f is a transfer function, usually a non-linear function. The following defaults to hardlim ()

t is the neuron output

Mathematical representation t = f (WA '+ b)

W is the weight vector

A is the input vector and A 'is the transpose of the A vector

b is the offset

f is the transfer function

It can be seen that the function of a neuron is to obtain the inner product of the input vector and the weight vector, and then obtain a scalar result through a non-linear transfer function.

The role of a single neuron: divide an n-dimensional vector space into two parts with a hyperplane (called the judgment boundary). Given an input vector, the neuron can determine which side of the hyperplane the vector is on.

The equation of this hyperplane: Wp + b = 0

W weight vector

b bias

Vector on p hyperplane

Basic characteristics of artificial neural networks

Artificial neural network is a nonlinear, adaptive information processing system composed of a large number of processing units interconnected. It is proposed on the basis of modern neuroscience research results, and attempts to process information by simulating the processing and memory of the neural network of the brain. Artificial neural networks have four basic characteristics:

(1) Non-linear Non-linear relationship is a universal characteristic in nature. Brain wisdom is a non-linear phenomenon. Artificial neurons are in two different states of activation or inhibition, and this behavior is mathematically non-linear

Artificial neural networks relationship. A network of threshold neurons has better performance and can improve fault tolerance and storage capacity.

(2) Non-limiting A neural network is usually made up of multiple neurons. The overall behavior of a system depends not only on the characteristics of individual neurons, but may also be mainly determined by the interactions and interconnections between units. Simulate the non-limiting nature of the brain through numerous connections between cells. Associative memory is a classic example of non-limitation.

(3) Very qualitative artificial neural network has the ability of self-adaptation, self-organization and self-learning. Not only can the information processed by neural networks have various changes, but also the nonlinear dynamic system itself is constantly changing while processing the information. Iterative processes are often used to describe the evolution of dynamic systems.

(4) Non-convexity The evolution direction of a system will depend on a certain state function under certain conditions. For example, the energy function, its extreme value corresponds to the relatively stable state of the system. Non-convexity means that this function has multiple extreme values, so the system has multiple stable equilibrium states, which will lead to the diversity of system evolution.

In the artificial neural network, the neuron processing unit can represent different objects, such as features, letters, concepts, or some meaningful abstract patterns. There are three types of processing units in the network: input units, output units, and hidden units. The input unit accepts signals and data from the outside world; the output unit realizes the output of system processing results; the hidden unit is between the input and output units and cannot

Artificial neural networks A unit viewed from outside the system. The connection weight between neurons reflects the connection strength between units, and the representation and processing of information is reflected in the connection relationship of network processing units. Artificial neural network is a kind of non-programmed, adaptive, and brain-style information processing. Its essence is to obtain a parallel and distributed information processing function through the transformation and dynamic behavior of the network, and imitate people at different levels and levels. Information processing functions of the cerebral nervous system. It is an interdisciplinary discipline that involves neuroscience, thinking science, artificial intelligence, computer science and other fields.

Artificial neural network is a parallel distributed system that uses a completely different mechanism from traditional artificial intelligence and information processing technology. It overcomes the defects of traditional artificial intelligence based on logical symbols in processing intuitive and unstructured information. Features of self-organizing and real-time learning. ^[1]

History of artificial neural networks

In 1943, the psychologist WSMcCulloch and the mathematical logician W. Pitts established neural networks and mathematical models, called MP models. They proposed a formal mathematical description of the neurons and a network structure method through the MP model, and proved that a single neuron can perform logical functions, thus initiating the era of artificial neural network research. In 1949, psychologists proposed a variable intensity of synaptic connections. In the 1960s, artificial neural networks were further developed, and more comprehensive neural network models were proposed.

Artificial neural networks , Including perceptrons and adaptive linear elements. M. Minsky et al. Carefully analyzed the functions and limitations of neural network systems represented by perceptrons, and published a book "Perceptron" in 1969, stating that perceptrons cannot solve high-order predicate problems. Their arguments greatly influenced the research of neural networks, coupled with the achievements of serial computers and artificial intelligence at that time, obscured the necessity and urgency of developing new approaches to new computers and artificial intelligence, and made artificial neural network research at a low tide . During this period, some researchers of artificial neural networks were still committed to this research, and proposed adaptive resonance theory (ART network), self-organizing mapping, cognitive machine network, and also studied the mathematical theory of neural networks. The above research has laid the foundation for the research and development of neural networks. In 1982, JJ Hopfield, a physicist at the California Institute of Technology in the United States, proposed the Hopfield neural grid model, introduced the concept of "calculated energy", and gave a judgment on network stability. In 1984, he also proposed the continuous-time Hopfield neural network model, which did pioneering work for the research of neural computers, pioneered a new way for neural networks to use in associative memory and optimize computing, and strongly promoted the research of neural networks. In 1985, another scholar proposed the Boltzmann model, and used statistical thermodynamic simulated annealing technology in the study to ensure that the entire system tended to a global stability point. In 1986, he studied the cognitive microstructure and proposed the theory of parallel distributed processing. In 1986, Rumelhart, Hinton, Williams developed the BP algorithm. Rumelhart and McClelland published "Parallel distribution processing: explorations in the microstructures of cognition." So far, the BP algorithm has been used to solve a large number of practical problems. In 1988, Linsker proposed a new self-organizing theory for perceptron networks, and formed the maximum mutual information theory based on Shanon's information theory, which ignited the light of NN-based information application theory. In 1988, Broomhead and Lowe proposed a hierarchical network design method using a radial basis function (RBF), thereby linking the design of the NN with numerical analysis and linear adaptive filtering. In the early 1990s, Vapnik et al. Proposed the concepts of support vector machines (SVM) and VC (Vapnik-Chervonenkis) dimension. The research on artificial neural networks has been valued by various developed countries. The United States Congress passed a resolution to designate the decade beginning on January 5, 1990 as the "decade of the brain." "Year" becomes a global act. In Japan's "Real World Computing (RWC)" project, research on artificial intelligence has become an important component.

Artificial neural network network model

The artificial neural network model mainly considers the topology of the network connection, the characteristics of the neurons, and the learning rules. At present, there are nearly 40 neural network models, including back propagation networks, perceptrons, self-organizing maps, Hopfield networks, Boltzmann machines, adaptive resonance theory, and so on. According to the connected topology, the neural network model can be divided into: ^[1]

Artificial neural networks

Artificial neural network forward network

Each neuron in the network accepts the input from the previous level and outputs it to the next level. There is no feedback in the network and it can be represented by a directed acyclic graph. This kind of network realizes the transformation of the signal from the input space to the output space, and its information processing ability comes from the multiple compounding of simple nonlinear functions. The network structure is simple and easy to implement. A backhaul network is a typical forward network. ^[2]

Artificial neural network feedback network

There is feedback between neurons in the network, which can be represented by an undirected complete graph. The information processing of this neural network is the transformation of state, which can be processed by dynamic system theory. The stability of the system is closely related to the associative memory function. Hopfield networks and Boltzmann machines belong to this type.

Artificial neural network learning types

Learning is an important part of neural network research, and its adaptability is achieved through learning. Adjust the weights according to the changes in the environment to improve the behavior of the system. The Hebb learning rules proposed by Hebb laid the foundation for the learning algorithm of neural networks. The Hebb rule states that the learning process finally occurs at the synapse sites between neurons, and the strength of the synapse connection changes with the activity of the neurons before and after the synapse. On this basis, various learning rules and algorithms have been proposed to meet the needs of different network models. Effective learning algorithms make God

Artificial neural networks The network can adjust the connection weights to construct an internal representation of the objective world, and form a unique information processing method. Information storage and processing are reflected in the connection of the network.

classification

According to different learning environments, neural network learning methods can be divided into supervised learning and unsupervised learning. In supervised learning, the training sample data is added to the network input, and the corresponding expected output is compared with the network output to obtain an error signal, which controls the adjustment of the weight connection strength and converges to one after multiple training Determined weight. When the sample situation changes, the weights can be modified to adapt to the new environment after learning. Neural network models using supervised learning include backpropagation networks, perceptrons, etc. In unsupervised learning, the standard sample is not given in advance, and the network is directly placed in the environment, and the learning phase and the working phase become one. At this time, the change of the learning law obeys the evolution equation of connection weights. The simplest example of unsupervised learning is Hebb learning rules. Competitive learning rules are a more complex example of unsupervised learning, which are weight adjustments based on established clusters. Self-organizing maps and adaptive resonance theory networks are typical models related to competitive learning.

Artificial neural network analysis method

Study the nonlinear dynamic properties of neural networks, mainly using dynamic system theory, nonlinear programming theory, and statistical theory to analyze the evolution of neural networks and the properties of attractors, explore the collaborative behavior and collective computing functions of neural networks, and understand Neural Information Processing Mechanism. In order to explore the possibility of neural network processing information in terms of integrity and ambiguity, the concepts and methods of chaos theory will come into play. Chaos is a mathematical concept that is quite difficult to define precisely. Generally speaking, "chaos" refers to the non-deterministic behavior shown in the dynamic system described by the deterministic equation, or it is called the deterministic randomness. "Determinism" is caused by internal reasons rather than external noise or interference, and "randomness" refers to its irregular and unpredictable behavior, which can only be described by statistical methods. Chaotic dynamic system

Artificial neural networks The main feature of the is its sensitive dependence on the initial conditions, and chaos reflects its inherent randomness. Chaos theory refers to the basic theories, concepts, and methods of describing nonlinear dynamic systems with chaotic behavior. It understands the complex behavior of dynamic systems as inherent in the process of material, energy, and information exchange with the outside world. Structural behavior, rather than foreign and accidental behavior, is a steady state. The stationary states of chaotic dynamic systems include: stationary, stationary, periodic, quasi-synchronous, and chaotic solutions. Chaotic trajectory is the result of a combination of overall stability and local instability, which is called a singular attractor. A singular attractor has the following characteristics: (1) a singular attractor is an attractor, but it is neither a fixed point nor a periodic solution; (2) a singular attractor is indivisible, that is, it cannot be divided into two And two or more attractors; (3) it is very sensitive to initial values, and different initial values can lead to very different behaviors. ^[3]

Features of artificial neural network

The characteristics and advantages of artificial neural networks are mainly manifested in three aspects:

First, it has a self-learning function. For example, when image recognition is implemented, only many different image templates and corresponding recognition results should be input into the artificial neural network first, and the network will slowly learn to recognize similar images through the self-learning function. The self-learning function is particularly important for prediction. It is expected that artificial neural network computers in the future will provide economic forecasts, market forecasts, and profit forecasts for human beings, and their application prospects are very promising.

Second, it has Lenovo memory function. This kind of association can be realized with the feedback network of artificial neural network.

Third, it has the ability to find optimal solutions at high speed. Finding an optimal solution to a complex problem often requires a large amount of calculation. Using a feedback-type artificial neural network designed for a certain problem and using the computer's high-speed computing capabilities, it may quickly find an optimal solution.

Research direction of artificial neural network

The research of neural network can be divided into two aspects: theoretical research and applied research.

Theoretical research can be divided into the following two categories: ^[4]

1. Use neurophysiology and cognitive science to study human thinking and intelligent mechanisms.

2.Using the research results of neural basic theory, using mathematical methods to explore more perfect functions and better performance

Artificial neural networks The more advanced neural network models, in-depth study of network algorithms and performance, such as: stability, convergence, fault tolerance, robustness, etc .; develop new network mathematical theory, such as: neural network dynamics, nonlinear neural fields, etc.

Applied research can be divided into the following two categories:

1. Research on software simulation and hardware implementation of neural network.

2. Research on the application of neural networks in various fields. These areas include

: Pattern recognition, signal processing, knowledge engineering, expert system, optimized combination, robot control, etc. With the continuous development of neural network theory and related theories and related technologies, the application of neural networks will definitely be deepened.

Development trend of artificial neural networks

The unique nonlinear adaptive information processing capability of artificial neural networks overcomes the shortcomings of traditional artificial intelligence methods in intuition, such as patterns, speech recognition, and unstructured information processing, and makes them suitable for neural expert systems, pattern recognition, intelligent control, Combination optimization, prediction and other fields have been successfully applied. The combination of artificial neural network and other traditional methods will promote the continuous development of artificial intelligence and information processing technology. In recent years, artificial neural networks are developing more deeply on the way to simulate human cognition. Combining them with fuzzy systems, genetic algorithms, and evolutionary mechanisms to form computational intelligence has become an important direction for artificial intelligence and will be developed in practical applications. . The application of information geometry to the study of artificial neural networks opens a new way for the theoretical research of artificial neural networks. Research on neural computers has developed rapidly, and existing products have entered the market. Optoelectronic combined neural computer provides good conditions for the development of artificial neural networks.

Neural networks have been well applied in many fields, but there are still many aspects to be studied. Among them, the combination of neural networks with other advantages such as distributed storage, parallel processing, self-learning, self-organization, and non-linear mapping, as well as the resulting hybrid methods and systems, has become a major research hotspot. Since other methods also have their own advantages, the neural network is combined with other methods to learn from each other's strengths, and then can obtain better application results. The current work in this area includes the fusion of neural networks and fuzzy logic, expert systems, genetic algorithms, wavelet analysis, chaos, rough set theory, fractal theory, evidence theory, and gray systems.

The following mainly analyzes the fusion of neural network and wavelet analysis, chaos, rough set theory and fractal theory.

Combination with wavelet analysis

In 1981, when French geologist Morlet sought geological data, he carried out creative research on the similarities and differences, characteristics, and function construction of Fourier transform and windowed Fourier transform. He first proposed the concept of "wavelet analysis" and established Named Morlet Wavelet. Since 1986, due to the foundation work of YMeyer, S. Mallat, and IDaubechies, wavelet analysis has rapidly developed into an emerging discipline. Meyer's "Wavelet and Operator" and Daubechies' "Wavelet Ten Lectures" are the most authoritative works in the field of wavelet research.

Wavelet transform is a breakthrough to Fourier analysis method. It not only has good localization properties in both time and frequency domains, but also has good resolution for low-frequency signals in the frequency domain and high-frequency signals in the time domain, so that it can gather any details of the object. Wavelet analysis is equivalent to a mathematical microscope with functions of zooming in, zooming out, and panning, and examining the dynamic characteristics of the signal by examining changes at different magnifications. Therefore, wavelet analysis has become a powerful tool in many fields such as geophysics, signal processing, image processing, and theoretical physics.

The wavelet neural network combines the good time-frequency localization characteristics of the wavelet transform and the self-learning function of the neural network, so it has strong approximation and fault tolerance. In the combination method, the wavelet function can be used as a basis function to construct a neural network to form a wavelet network, or the wavelet transform can be used as an input pre-processing tool of a feedforward neural network, that is, the multi-resolution characteristics of the wavelet transform are used to process the process state signal. The signal-to-noise separation is realized, and the state characteristics that have the greatest influence on processing errors are extracted as the input of the neural network.

The wavelet neural network has applications in motor fault diagnosis, high-voltage grid fault signal processing and protection research, bearing and other mechanical fault diagnosis, and many other aspects. The wavelet neural network is used for intelligent control of induction servo motors, so that the system has good tracking control. Performance and good robustness, using wavelet packet neural network for intelligent diagnosis of cardiovascular disease, wavelet layer for adaptive feature extraction in time-frequency domain, and forward neural network for classification, the correct classification rate reaches 94%.

Although wavelet neural network is applied in many aspects, there are still some shortcomings. Starting from the requirements of extraction accuracy and wavelet transform real-time, it is necessary to construct some special wavelet bases according to the actual situation in order to obtain better results in the application. In addition, real-time requirements in applications also need to be combined with the development of DSP to develop special processing chips to meet this requirement.

Chaotic neural network

The first definition of chaos was first proposed by Li-Yorke in the 1970s. Because it has a wide range of application values, it has received widespread attention from all sides since its appearance. Chaos is a kind of random movement in a certain system. Chaos is a more common phenomenon existing in non-linear systems. Chaotic motion has the characteristics of ergodicity and randomness. Iterate through all states without repeating itself. Chaos theory determines non-linear dynamic chaos. The purpose is to reveal the simple laws that may be hidden behind seemingly random phenomena, in order to find common laws that are generally followed by a large class of complex problems.

In 1990, Kaihara, T. Takabe, and M. Toyoda, etc. first proposed a chaotic neural network model based on the chaotic characteristics of biological neurons, and introduced chaos into the neural network, making the artificial neural network have chaotic behavior and closer to the actual human brain nerves. Therefore, chaotic neural network is considered as one of the intelligent information processing systems that can realize its real-world computing, and it has become one of the main research directions of neural networks.

Compared with conventional discrete Hopfield neural networks, chaotic neural networks have richer non-linear dynamic characteristics, mainly as follows: the introduction of chaotic dynamic behavior in neural networks; the synchronization characteristics of chaotic neural networks; the attraction of chaotic neural networks child.

When the neural network is used in practical applications, the network input has a large variation, the inherent fault tolerance of the application network is often insufficient, and amnesia often occurs. Dynamic memory of chaotic neural network belongs to deterministic dynamic movement. Memory occurs on the trajectory of chaotic attractors. Through continuous movement (recall process), one by one thinks of memory patterns, especially for those close to the state-space distribution or partially overlapping. In the memory mode, chaotic neural networks can always be reproduced and identified through dynamic associative memory without confusion. This is a characteristic of chaotic neural networks that will greatly improve the memory capacity of Hopfield neural networks. The existence of the attraction domain of the chaotic attractor forms the inherent fault tolerance function of the chaotic neural network. This will play an important role in engineering applications such as complex pattern recognition and image processing.

Another reason that chaotic neural networks attract attention is that chaos exists in real neurons and neural networks of living organisms and plays a certain role. This has been confirmed by zoophysical electrophysiological experiments.

Chaotic neural networks have received great attention in the fields of dynamic associative memory, system optimization, information processing, and artificial intelligence due to their complex dynamic characteristics. For the chaotic neural network has associative memory function, but its searching process is unstable, a control method is proposed to control the chaos phenomenon in the chaotic neural network. The application of chaotic neural network in combinatorial optimization is studied.

In order to better apply the dynamic characteristics of the chaotic neural network and effectively control the chaos phenomenon, it is still necessary to further improve and adjust the structure of the chaotic neural network and further study the algorithm of the chaotic neural network.

Based on rough set theory

Rough Sets theory was first proposed by Z.Pawlak, a professor at the Warsaw University of Technology in Poland in 1982. It is a mathematical theory for analyzing data, studying incomplete data, expression of inaccurate knowledge, learning, induction and other methods. Rough set theory is a new mathematical tool for fuzzy and uncertain knowledge. Its main idea is to derive the decision or classification rules of the problem through knowledge reduction under the premise of keeping the classification ability unchanged. At present, rough set theory has been successfully applied in the fields of machine learning, decision analysis, process control, pattern recognition and data mining.

The common point of rough sets and neural networks is that they both work well in natural environments. However, the rough set theory method simulates human abstract logical thinking, and the neural network method simulates image intuitive thinking, so the two have different characteristics. Rough set theory takes qualitative, quantitative, or mixed information that is closer to the way people describe things as input. The mapping relationship between input space and output space is simplified through a simple decision table. It considers the differences in knowledge expression. The importance of attributes determines which knowledge is redundant and which knowledge is useful. Neural networks use the idea of non-linear mapping and parallel processing methods to use the neural network structure to express the implicit function encoding of input and output association knowledge.

There are two major differences between the rough set theory method and the neural network method for processing information. One is that neural network processing information generally cannot simplify the dimension of the input information space. When the dimension of the input information space is large, Not only is the network complex in structure, but also the training time is long; but the rough set method can not only remove redundant input information, but also simplify the expression space dimension of the input information by discovering the relationship between the data. The second is that the rough set method is more sensitive to noise in the processing of practical problems, so the results of learning inference using noiseless training samples are not effective in noisy environments. The neural network method has a better ability to suppress noise interference.

Therefore, the two are combined, and the information is pre-processed using the rough set method, that is, the rough set network is used as the pre-system, and then the pre-processed information structure of the rough set method is used to form a neural network information processing system. Through the combination of the two, not only can reduce the number of attributes of information expression, reduce the complexity of the neural network composition system, but also have strong fault tolerance and anti-interference ability, which provides a powerful method for processing uncertain and incomplete information. way.

At present, the combination of rough sets and neural networks has been used in the fields of speech recognition, expert systems, data mining, and fault diagnosis. Neural networks and rough sets are used for automatic recognition of sound source locations, and neural networks and rough sets are used in expert systems. In the acquisition of knowledge, better results are achieved than traditional expert systems, in which rough sets deal with uncertain and imprecise data, and neural networks perform classification work.

Although the combination of rough sets and neural networks has been used in many fields of research, in order to make this method play a greater role, the following issues need to be considered: the rough set theory method to simulate human abstract logical thinking and the neural network to simulate intuitive thinking The methods are more effectively combined; the development of the integrated software and hardware platforms of the two improves their practicability.

Combination with fractal theory

Since Benoit B. Mandelbrot, a professor of mathematics at Harvard University in the United States, introduced the concept of fractal in the mid-1970s, fractal geometry has developed into a scientific methodology--fractal theory, and is hailed as the pioneer of 20th century mathematics Important stage. Now it has been widely used in almost all fields of natural sciences and social sciences, and has become one of the frontier research topics of many disciplines in the world today.

Due to the rapid development in many disciplines, fractal has become a discipline that describes the regularity of many irregular things in nature. It has been widely used in various fields such as biology, geography, astronomy, computer graphics and so on.

Using fractal theory to explain the phenomena of irregularity, instability and highly complex structures in nature can receive significant results. Combining neural networks with fractal theory makes full use of neural network non-linear mapping, computing power, Adaptation and other advantages can achieve better results.

The application fields of fractal neural network are image recognition, image coding, image compression, and fault diagnosis of mechanical equipment systems. The fractal image compression / decompression method has the advantages of high compression rate and low loss rate, but the computing power is not strong. Because the neural network has the characteristics of parallel operation, using the neural network in the fractal image compression / decompression improves the original The computing power of the method. The neural network and fractal are used for fruit shape recognition. First, the fractal is used to obtain the irregularity of several kinds of fruit contour data. Then the three-layer neural network is used to identify these data, and then the irregularity is evaluated.

Fractal neural networks have achieved many applications, but there are still some issues worthy of further study: the physical meaning of fractal dimensions; computer simulation and practical application research of fractals. With the continuous deepening of research, the fractal neural network will be constantly improved and achieve better application results. ?

Application analysis of artificial neural network

After decades of development, neural network theory has achieved widespread success in many research areas such as pattern recognition, automatic control, signal processing, aided decision making, and artificial intelligence. The application status of neural networks in some fields is introduced below. ^[3]

Application of artificial neural network in the field of information

In dealing with many problems, the information source is incomplete and contains artifacts. Decision rules are sometimes contradictory and sometimes out of order. This brings great difficulties to traditional information processing methods, but neural networks can be very good. To deal with these problems, and give reasonable identification and judgment.

Information processing

The problems to be solved by modern information processing are very complicated. Artificial neural networks have the functions of imitating or replacing human thinking. They can realize automatic diagnosis, problem solving, and solve problems that traditional methods cannot or cannot solve. The artificial neural network system has high fault tolerance, robustness and self-organization. Even if the connection line is damaged to a high degree, it can still be in an optimal working state. This is widely used in military system electronic equipment. application. The existing intelligent information systems include intelligent instruments, automatic tracking and monitoring instrument systems, automatic control guidance systems, automatic fault diagnosis and alarm systems, and the like.

Pattern recognition

Pattern recognition is the process of describing and identifying, classifying, and interpreting various forms of information that characterize a thing or phenomenon. This technology is based on Bayesian probability theory and Shennong's information theory. The processing of information is closer to the logical thinking process of the human brain. There are currently two basic pattern recognition methods, namely statistical pattern recognition method and structural pattern recognition method. Artificial neural network is a commonly used method in pattern recognition. Artificial neural network pattern recognition methods developed in recent years gradually replace traditional pattern recognition methods. After years of research and development, pattern recognition has become the current relatively advanced technology and is widely used in text recognition, speech recognition, fingerprint recognition, remote sensing image recognition, face recognition, handwritten character recognition, industrial fault detection, precise guidance, etc. aspect.

Application of artificial neural network in medicine

Due to the complexity and unpredictability of the human body and diseases, the biological signals and information are expressed in terms of their expressions and changes (self changes and changes after medical intervention), which are detected and signaled, and the acquired data and information are analyzed. There are many complicated non-linear relations in many aspects such as decision-making, which are suitable for the application of artificial neural network. The current research covers almost every aspect from basic medicine to clinical medicine, mainly applied to the detection and automatic analysis of biological signals, medical expert systems, etc.

Detection and analysis of biological signals

Most medical testing equipment outputs data in the form of continuous waveforms, which are the basis for diagnosis. Artificial neural network is an adaptive dynamic system connected by a large number of simple processing units. It has huge parallelism, distributed storage, self-organization of adaptive learning and other functions. It can be used to solve biomedical signal analysis and processing. Problems that are difficult or impossible to solve with conventional law. The application of neural networks in the detection and processing of biomedical signals mainly focuses on the analysis of EEG signals, the extraction of auditory evoked potential signals, the recognition of signals such as myoelectricity and gastrointestinal electricity, the compression of ECG signals, and the recognition of medical images And processing.

Medical expert system

The traditional expert system is to store the experience and knowledge of experts in a computer in the form of rules, establish a knowledge base, and conduct medical diagnosis by logical reasoning. However, in actual applications, as the size of the database increases, knowledge will "explode", and there will also be "bottlenecks" in the way of knowledge acquisition, resulting in very low work efficiency. The research on expert systems based on nonlinear parallel processing neural networks points out new development directions, solves the above problems of expert systems, and improves the ability of knowledge reasoning, self-organization, and self-learning. It has been widely used and developed in the system. In the research of related fields such as anesthesia and critical medicine, the analysis and prediction of multiple physiological variables are involved. In the clinical data, there are some undiscovered or inconclusive relationships and phenomena, signal processing, and automatic discrimination detection of interference signals The prediction of various clinical conditions can be applied to artificial neural network technology.

Application of artificial neural network in economic field

Market price forecast

The analysis of changes in commodity prices can be summarized as a comprehensive analysis of many factors affecting the relationship between market supply and demand. Traditional statistical economics methods are difficult to make scientific predictions of price changes due to their inherent limitations. Artificial neural networks are easy to process incomplete, fuzzy, uncertain, or inconsistent data. Price prediction has advantages that cannot be compared with traditional methods. Starting from the market price determination mechanism, a more accurate and reliable model is established based on complex and changing factors such as the number of households, per capita disposable income, loan interest rates, and urbanization levels that affect commodity prices. This model can scientifically predict the trend of commodity prices and obtain accurate and objective evaluation results.

Risk assessment

Risk refers to the possibility of economic or financial loss, natural damage or damage caused by the uncertainty in the process of engaging in a specific activity. The best way to prevent risks is to make scientific predictions and assessments of risks in advance. The artificial neural network prediction idea is to construct the structure and algorithm of the credit risk model suitable for the actual situation according to the actual risk source, obtain the risk evaluation coefficient, and then determine the solution to the actual problem. Empirical analysis using this model can make up for the shortcomings of subjective assessment and achieve satisfactory results.

Application of artificial neural network in the field of control

Artificial neural networks have been widely used in control systems due to their unique model structure and inherent nonlinear simulation capabilities, as well as outstanding features such as high self-adaptation and fault tolerance. Based on the framework structure of various controllers, it adds a non-linear adaptive learning mechanism, so that the controller has better performance. The basic control structures include supervisory control, direct inverse model control, model reference control, internal model control, predictive control, optimal decision control, and so on.

Application of artificial neural network in the field of transportation

This year, people have begun to study the application of neural networks in transportation systems. The transportation problem is highly non-linear, and the available data is usually large and complex. Processing related problems with neural networks has its great advantages. Applications include vehicle driver behavior simulation, parameter estimation, road maintenance, vehicle detection and classification, traffic mode analysis, cargo operation management, traffic flow prediction, transportation strategy and economy, transportation environmental protection, air transportation, automatic navigation of ships, and Good results have been achieved in the areas of ship identification, subway operation, and traffic control.

Application of artificial neural network in the field of psychology

Since the formation of the neural network model, it has been inseparably linked to psychology. Neural networks are abstracted from the information processing functions of neurons, and the training of neural networks reflects cognitive processes such as sensory, memory, and learning. Through continuous research, people have changed the structural model and learning rules of artificial neural networks, and explored the cognitive functions of neural networks from different perspectives, laying a solid foundation for their research in psychology. In recent years, artificial neural network models have become an indispensable tool for exploring advanced psychological process mechanisms such as social cognition, memory, and learning. The artificial neural network model can also study the cognitive deficits of patients with brain injury, which challenges the traditional cognitive positioning mechanism.

Although artificial neural networks have made some progress, there are still many shortcomings, such as: the application is not wide enough and the results are not accurate enough; the training speed of existing model algorithms is not high enough; the degree of integration of the algorithms is not high enough; at the same time we hope that Find a new breakthrough point in theory and establish a new general model and algorithm. Further research on biological neuron systems is needed to continuously enrich people's understanding of human brain nerves.