Artificial Neural Network Evaluation Method
Artificial neuron is the basic processing unit of artificial neural network, and an important part of artificial intelligence is artificial neural network. The artificial neural network is a mathematical model that simulates the biological neuron system, and the information is mainly received through neurons.
First, the artificial neuron uses the connection strength to amplify the generated signal; then, it receives the weighted accumulation of the output of all neurons connected to it; finally, compares the neuron with the weighted sum one by one, and activates when it is greater than the threshold With artificial neurons, signals are sent to neurons in the layer above it is connected, but not vice versa.
An important model of the artificial neural network is the Back-Propagation Model (BP model for short).
For a backpropagation network with n input nodes and m output nodes, the relationship between input and output can be regarded as a mapping from n-dimensional space to m-dimensional space. Due to the large number of nonlinear nodes in the network, it can be highly nonlinear.
(1) The steps of neural network evaluation method The purpose of using neural network to evaluate reclamation potential is to produce an expected evaluation result for the input of a certain index. In the process, the weight of the connection arc of the network needs to be continuously adjusted. . (1) Initialize the weights of all connected arcs.
In order to ensure that the network will not be saturated and abnormal, it is generally set to a small random number. (2) Input a set of training data into the network, and calculate the output value of the network.
(3) Calculate the deviation between the expected value and the output value, and then reversely calculate from the output layer to the first hidden layer, and adjust the weight of each arc to make it develop in the direction of reducing the deviation.
(4) Repeat the above steps and repeatedly calculate each group of training data in the training set until the deviation between the two reaches a level that can be recognized. (2) Establishment of artificial neural network model (1) Determine the number of input layers.
According to the actual situation of the evaluation object, the number of input layers is the number of selected evaluation indicators. (2) Determine the number of hidden layers.
Usually the most ideal neural network has only one hidden layer. The input signal can be separated by hidden nodes and then combined into a new vector. Its operation is fast, which can simplify complex things and reduce unnecessary troubles. (3) Determine the number of hidden layer nodes.
According to the empirical formula: In the disaster-damaged land reclamation formula: j—the number of hidden layers; n—the number of input layers; m—the number of output layers. The structure of the artificial neural network model is shown in Figure 5-2.
Figure 5-2 Structure diagram of artificial neural network (according to Zhou Lihui, 2004) (3) The calculation of artificial neural network inputs the index information (X1, X2, X3, ..., Xn) of the evaluated object, and calculates the actual output value Yj.
The reclamation of disaster-damaged land compares the known output with the calculated output, and modifies the weights and thresholds of K-layer nodes. In the disaster-damaged land reclamation formula: wij——connection weight and threshold of node j in K-1 layer; η——coefficient (0<η<1); Xi——output of node i.
Output result: Cj=yj(1-yj)(dj-yj) (5-21) where: yj—the actual output value of node j; dj—the expected output value of node j.
Because it is impossible to compare the output of hidden nodes, it can be deduced: In the disaster-damaged land reclamation formula: Xj—the actual output value of node j.
It is a process of taking turns to replace, each iteration will adjust the W value, so after repeated replacement, it will not stop until the deviation between the calculated output value and the expected output value is within the allowable value range.
Using the artificial neural network method to evaluate the reclamation potential is actually to establish the mapping relationship between land reclamation impact evaluation factors and reclamation potential.
As long as the selected network structure is appropriate, the above-mentioned mapping relationship can be infinitely approached by using the approximation of the artificial neural network function, so it is appropriate to use the artificial neural network method to evaluate the reclamation potential of disaster-damaged land.
(4) Advantages and disadvantages of the artificial neural network method Compared with other methods, the artificial neural network method has the following advantages: (1) It uses the optimal training principle to perform repeated calculations, and continuously debugs the neural network structure until a relatively stable the result of.
Therefore, adopting this method to evaluate reclamation potential can eliminate many subjective factors and ensure the authenticity and objectivity of the evaluation results of reclamation potential. (2) The error of the obtained evaluation result is relatively small, and the system error can be reduced through repeated iterations, which can meet any accuracy requirements.
(3) Good dynamics, by increasing the number of reference samples and over time, dynamic tracking comparison and deeper learning can be realized.
(4) It is based on nonlinear functions, which is closer to the complex nonlinear dynamic economic system, and can reflect the reclamation potential of disaster-damaged land more truly and accurately, and is more applicable than traditional evaluation methods.
But the artificial neural network also has certain deficiencies: (1) The artificial neural network algorithm adopts an optimization algorithm, which continuously adjusts the weights between the connected neurons through iterative calculations until the global optimization is achieved.
However, the error surface is quite complicated, and the neural network will fall into a local minimum point if you are not careful during the calculation process.
(2) The error propagates backwards through the output layer. The more hidden layers are, the more inaccurate the reverse propagation deviation is when it is close to the input layer, and the evaluation efficiency is also affected to a certain extent, and the convergence speed is not timely. It caused deviation in the evaluation results of reclamation potential in individual areas.
Google AI Writing Project: Neural Network Pseudo-Original
Regarding the selection of indicators when using bp neural network for classification 10
Neural Network Binary Classification vs Multi Classification
Two classifications: the labels are 0 and 1, and the result of the network output is processed by the sigmoid activation function. The value range of the output value is between 0 and 1. If it is less than 0.5, it is regarded as label 0, and if it is greater than or equal to 0.5, it is label 1. Multi-classification: The label has multiple values, and the network needs to output a z-score vector whose dimension is consistent with the number of labels. The vector needs to be activated by softmax and converted into the probability corresponding to each label (probability sum is 1), and the determined label is the highest probability that.
How to define qualified and basic qualified in bp neural network
You mean classification. How to classify depends on how the training samples are divided. The output of possible samples is quantified as follows: qualified 1, basically qualified 0.5, unqualified 0. Or qualified 100, basically qualified 010, unqualified 001. The neural network can be used for classification, clustering, prediction, etc.
The neural network needs a certain amount of historical data. Through the training of historical data, the network can learn the hidden knowledge in the data. In your problem, you must first find some characteristics of certain problems and the corresponding evaluation data, and use these data to train the neural network.
Neural Network Classification Problem
Neural networks are a buzzword in the field of new technologies. Many people have heard the term, but few really understand what it is. The purpose of this article is to introduce all the basics about neural networks including their functions, general structure, related terms, types and their applications.
The word "neural network" actually comes from biology, and the correct name for what we mean by neural networks should be "artificial neural networks (ANNs)". In this article, I will use both terms interchangeably.
A true neural network is made up of anywhere from a few to billions of cells called neurons (the tiny cells that make up our brains), which are connected in different ways to form a network. Artificial neural networks are attempts to mimic this biological architecture and its operation.
Here's the rub: we don't know much about neural networks in biology! Therefore, neural network architectures vary greatly between different types, and all we know is the basic structure of neurons.
The neuron ------------------------------------------------ -------------------------------- Although it has been confirmed that there are approximately 50 to 500 different types of neurons in our brain , but most of them are specialized cells based on basic neurons.
Basic neurons include synapses, soma, axon and dendrites.
Synapses are responsible for the connections between neurons, they are not directly physically connected, but there is a small gap between them that allows electrical signals to jump from one neuron to another.
These electronic signals are then handed over to the soma for processing and the processing results are passed to axon with its internal electronic signals. And axon will distribute these signals to dendrites.
Finally, the dendrites take these signals and pass them on to other synapses, and the next cycle continues. Like basic neurons in biology, artificial neural networks also have basic neurons.
Each neuron has a specific number of inputs, and weights are also set for each neuron. The weight is an indicator of the importance of the input data.
The neuron then computes a net value, which is the sum of all inputs multiplied by their weights. Each neuron has its own threshold (threshold), and when the total weight value is greater than the threshold, the neuron will output 1.
Otherwise, 0 is output. Finally, the output is sent to other neurons connected to this neuron to continue the remaining calculations.
Learning -------------------------------------------------- ------------------------------- As written above, the core of the problem is how to set the weight and threshold Woolen cloth?
There are as many different training styles in the world as there are network types. But some of the more well-known ones include back-propagation, delta rule and Kohonen training mode.
Due to the difference in the structure system, the training rules are also different, but most of the rules can be divided into two categories - supervised and unsupervised. Supervision-style training rules require a "teacher" to tell them what output a given input should produce.
The training rules then adjust all the required weight values (which is very complex in the network), and the whole process starts from scratch until the data can be correctly analyzed by the network. The training mode of supervision method includes back-propagation and delta rule.
Rules in a non-supervisory manner do not require a teacher, as the output they produce is further evaluated.
Architecture ------------------------------------------------- ------------------------------- In neural networks, the word obeying explicit rules is the most "ambiguous" .
Because there are so many different kinds of networks, from simple Boolean networks (Perceptrons), to complex self-adjusting networks (Kohonen), to thermodynamic network models (Boltzmann machines)!
And these all comply with the standard of a network architecture. A network consists of multiple "layers" of neurons, an input layer, a hidden layer, and an output layer. The input layer is responsible for receiving input and distributing it to the hidden layer (because the user cannot see these layers, so see as hidden layer).
These hidden layers are responsible for the required calculations and output the results to the output layer, and the user can see the final result. Now, to avoid confusion, I won't delve into the topic of architecture in more depth here.
More details on the different neural networks can be found in the Generation5 essays Although we have discussed neurons, training and architecture, it is not yet clear what the neural network actually does.
The Function of ANNs ---------------------------------------------- ---------------------------------- Neural networks are designed to work with patterns - they can be grouped into categories formula or associative.
Categorical networks can take a set of numbers and classify them. For example, the ONR program accepts an image of a number and outputs the number. Or the PPDA32 program accepts a coordinate and classifies it as class A or class B (the class is determined by the training provided).
For more practical uses, see the military radar in Applications in the Military, which can distinguish vehicles or trees. Associative patterns accept one set of numbers and output another set.
For example a HIR program accepts a 'dirty' image and outputs the closest image it has learned. Associative mode can be applied to complex applications such as signature, face, fingerprint recognition, etc.
The Ups and Downs of Neural Networks ---------------------------------------------- ------------------------------------- Neural networks have many advantages in this field, making it more more and more popular.
It's really good at type classification/recognition. Neural networks can handle exceptions and abnormal input data, which is important for many systems (such as radar and acoustic positioning systems). Many neural networks mimic biological neural networks, that is, they work after the way the brain works.
Neural networks are also aided by the development of neuroscience, allowing them to identify objects as accurately as humans and at the speed of computers! The future is bright, but for now... yes, there are some bad things about neural networks. This is usually due to lack of sufficiently powerful hardware.
The power of neural networks comes from processing information in parallel, that is, processing multiple pieces of data at the same time. Therefore, it is very time-consuming to simulate parallel processing on a serial machine.
Another problem with neural networks is that there are not enough defined conditions for building a network for a certain problem - there are too many factors to consider: training algorithm, architecture, number of neurons per layer, how many layers, data performance, etc., There are many more factors.
Therefore, as time becomes more and more important, most companies cannot afford to repeatedly develop neural networks to effectively solve problems.
Conclusion ------------------------------------------------- ------------------------------- Hope you can have a basic understanding of neural networks through this article.
Generation5 now has a lot of information about neural networks available, including articles and programs. We have examples of Hopfield, perceptrons (2) networks, and some back-propagation case studies.
Glossary ------------------------------------------------- ------------------------------- NN neural network, Neural Network ANNs artificial neural network, Artificial Neural Networks neurons neuron synapses neural Keys self-organizing networks self-adjusting networks networks modeling thermodynamic properties thermodynamic properties of the network model.
Artificial Neural Network Classification Method
Since the late 1980s, artificial neural network methods have been applied to the automatic classification of remote sensing images.
At present, in the automatic classification of remote sensing images, there are mainly the following artificial neural network methods that are widely used and studied: (1) BP (Back Propagation) neural network, which is a widely used feed-forward network. It belongs to a supervised classification algorithm. It integrates prior knowledge into network learning and makes maximum use of it. It has good adaptability and can obtain quite high accuracy in the case of a small number of categories. However, its network learning mainly uses error To modify the algorithm, when there are many types of recognition objects, with the expansion of the network scale, the required calculation process is longer, the convergence is slow and unstable, and the recognition accuracy is difficult to meet the requirements.
(2) Hopfield neural network. It is a feedback network. The Hebb rule is mainly used for learning, and the convergence speed of the calculation is generally faster.
This network was first proposed by American physicist JJ Hopfield in 1982, and it is mainly used to simulate the memory mechanism of biological neural networks.
The evolution process of Hopfield neural network state is a nonlinear dynamic system, which can be described by a set of nonlinear difference equations.
The stability of the system can be analyzed by the so-called "energy function". Under certain conditions, the energy of a certain "energy function" decreases continuously during the operation of the network, and finally tends to a stable equilibrium state.
The evolution process of Hopfield network is a process of computing associative memory or solving optimization problems. (3) Kohonen network.
This is a self-organizing neural network proposed by Kohonen (1981), a neural network expert at the University of Helsinki in Finland, which uses a learning algorithm without tutor information. Tasks like environment learning, automatic classification and clustering.
Its biggest advantage is that there is a similar relationship between the final adjacent clusters. Even if the sample is mapped to a wrong node during identification, it tends to be identified as the same factor or a similar factor. Very close to human identification characteristics.