It is expected that the future artificial neural network computer will provide economic prediction, market prediction and benefit prediction for human beings, and its application prospect is very bright. (2) has the function of associative storage. This association can be realized by using the feedback network of artificial neural network. (3) it has the ability to find the optimal solution at high speed.

Finding the optimal solution of a complex problem often requires a large amount of computation. Using a feedback artificial neural network designed for a problem and giving full play to the high-speed computing ability of the computer, the optimal solution may be found quickly.

Extended data: the unique nonlinear adaptive information processing ability of artificial neural network overcomes the shortcomings of traditional artificial intelligence methods in intuition, such as pattern, speech recognition, unstructured information processing. It has been successfully applied in the fields of neural expert system, pattern recognition, intelligent control, combinatorial optimization, prediction and so on.

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 network is developing more deeply on the road of simulating human cognition, which is combined with fuzzy system, genetic algorithm and evolutionary mechanism to form computational intelligence, which has become an important direction of artificial intelligence. it will be developed in practical application.

The application of information geometry to the study of artificial neural network opens up a new way for the theoretical research of artificial neural network. The research of neural computer has developed rapidly, and some products have entered the market. The neural computer with optoelectronic combination provides good conditions for the development of artificial neural network.

Neural network has been well applied in many fields, but there are still many aspects that need to be studied.

Among them, the combination of neural networks and other technologies with the advantages of distributed storage, parallel processing, self-learning, self-organization and nonlinear mapping, as well as the resulting hybrid methods and hybrid systems, has become a research hotspot.

As other methods also have their own advantages, so the combination of neural network and other methods, learn from each other, and then can obtain better application results.

At present, the work in this field includes the integration of neural network and fuzzy logic, expert system, genetic algorithm, wavelet analysis, chaos, rough set theory, fractal theory, evidence theory and grey system. Reference: Baidu encyclopedia-artificial neural network.

Through learning to find the rules hidden behind the data, and for the data outside the learning set with the same law, this trained network can give appropriate output, this ability is called generalization ability.

For the neural network, the more complex the neural network is, the higher the complexity of the neural network is, the greater the complexity capacity of the neural network is, of course, the better, and certainly not absolute, but it shows the problem of the capacity of a container. at this time, the generalization ability of the neural network is also stronger.

We need to know that structural complexity and sample complexity, sample quality, initial weight, learning time and other factors will affect the generalization ability of the neural network.

In order to ensure that the neural network has strong generalization ability, people have done a lot of research and obtained many generalization methods, including pruning algorithm, construction algorithm and evolutionary algorithm and so on. The generalization ability of artificial neural network is mainly due to the efficient feature set can be derived from the training set through unsupervised pre-learning.

Once complex problems are transformed into forms expressed by these features, they naturally become simpler. Conceptually, this is a bit like doing an intelligent coordinate transformation for the training set.

For example, if the training set is a lot of pictures of faces, then if the pre-training is done well, you can derive features such as nose, eyes, mouth, various basic face shapes, and so on. If the classification was done with these features instead of based on pixels, the result would be much better.

Although large neural networks have many parameters, they are less likely to be over-fitted because they are actually based on a small number of features.

At the same time, in view of the shortcomings that the neural network is easy to fall into local extremum, difficult to determine its structure and poor generalization ability, a support vector regression machine which can well solve the problems of small samples, nonlinearity and high dimension is introduced to predict the development index of oil and gas field.

The limitation of the neural network algorithm is that the mean function can be used, but this function will take the average value of the embedded and assign it as a new embedding. However, it is easy to see that for some different graphs, they will give the same embedding, so the mean function is not monojective.

The universality of the neural network algorithm is that it is usually easier to gain insight into the model because of the limitations of the model. After all, the knowledge about specific features that the network cannot learn is independent of the training process in application.

Training a graph classifier needs to identify what constitutes a class, that is, to find the attributes shared by the graph in the same class rather than in other classes, and then determine whether the new graph complies with the learned attributes.

However, if the above decision problem can be proved to be impossible by a graph neural network of a certain depth (and the test set is diversified enough), then we can be sure that the same network will not learn how to classify the test set correctly. this has nothing to do with what learning algorithm is used.

Therefore, when conducting experiments, we should focus on the network that is deeper than the lower limit.

In fact, the encyclopedia is introduced in detail, such as "artificial neural network is a thinking function that simulates the structure of the human brain, with strong self-learning and association functions, less manual intervention, higher accuracy, and less use of expert knowledge."

But the disadvantage is that it can not deal with and describe fuzzy information, can not make good use of the existing experience and knowledge, especially the learning and problem solving has black box characteristics, its work is not explainable, and it has high requirements for samples. Compared with the neural network, the fuzzy system has the advantages of easy to understand reasoning process, better utilization of expert knowledge and lower requirements for samples, but it also has some shortcomings such as too much manual intervention, slow reasoning speed and low precision. It is difficult to realize the function of adaptive learning, and how to generate and adjust membership functions and fuzzy rules automatically is also a thorny problem.

The main advantages of deep learning are as follows: 1: strong learning ability and strong learning ability. 2: the neural network with wide coverage, good adaptability and deep learning has many layers and wide breadth, and can be mapped to any function in theory, so it can solve very complex problems.

In order to achieve good accuracy, it needs the support of big data. The complexity of the graph model in deep learning leads to a sharp increase in the time complexity of the algorithm. in order to ensure the real-time performance of the algorithm, higher parallel programming skills and more and better hardware support are needed.

Therefore, only some scientific research institutions or enterprises with strong economic strength can use deep learning to do some cutting-edge and practical applications.

This makes it especially suitable for solving problems with complex internal mechanisms; the network can automatically extract "reasonable" solving rules by learning a set of examples with correct answers, that is, it has the ability of self-learning; the network has a certain ability of generalization and generalization.

It is difficult to solve the contradiction between the instance scale and the network scale of the application problem. This involves the relationship between the possibility and feasibility of network capacity, that is, learning complexity; the choice of network structure does not have a unified and complete theoretical guidance, and can only be selected by experience.

For this reason, some people call the structure selection of neural network an art. The structure of the network directly affects the approximation ability and generalization property of the network.

Therefore, how to choose the appropriate network structure in the application is an important problem; the newly added samples will affect the successfully learned network, and the number of features depicted by each input sample must be the same; the contradiction between the prediction ability (also known as generalization ability, generalization ability) and training ability (also known as approximation ability, learning ability) of the network.

In general, when the training ability is poor, the prediction ability is also poor, and to a certain extent, with the improvement of training ability, the prediction ability is also improved. However, this trend has a limit, when it reaches this limit, with the improvement of training ability, the prediction ability decreases, that is, the so-called "over-fitting" phenomenon occurs.

Probabilistic neural network. Deep neural network means that Microsoft has launched a new speech recognition software, which works by imitating the way the human brain thinks, so that the speech recognition speed of the software is faster and the recognition accuracy is higher.

Organic chemistry is a branch of chemistry, which studies the structure, properties and reactions of hydrocarbons and their derivatives. The study of the structure includes the study of the molecular structure and crystal structure of the compounds by means of spectroscopy, chemical calculation and computer simulation.

The study of properties includes the physical and chemical properties of compounds and the prediction of chemical reactivity.