How artificial neural networks work (Topic)

World Of Topics » Science » How artificial neural networks work

How artificial neural networks work


Almost every week you can hear about the next success of artificial intelligence, it beats the leading poker players, looks for employees for various companies and diagnoses skin cancer.

Yes, yes, this is the very thing that can find your page in social networks from a photo.

How neural networks work

The human brain is not able to quickly process large volumes of mathematical operations, even such simple devices as, for example, a calculator are easily ahead of a person, but unlike a calculating machine, it can quickly adapt to new conditions.

A person, for example, can easily recognize his interlocutor, despite the fact that he is in a noisy room and recognize a celebrity, even if she is wearing a mask with a mustache.

It's quite difficult for a computer to do this. Human thinking is the result of chemical processes in a network of neurons in his brain, which exchange signals with each other using electrical impulses.

Such collective activity contributes to the birth of perceptions and thoughts in the human mind. The interactions of neural networks, modeled with a computer, are an artificial neural network. The first mathematical model of an artificial neuron was known back in 1943, it was proposed by Warren McCulloch and Walter Pitts, but the possibility of their active use appeared only at the end of the twentieth century.

How does a neural network work?

The principle of operation of such a network is as follows. Signals are sent to the inputs of neurons, which are summed up, taking into account the weight, that is, the significance of each input. Further, the output signals of some neurons are fed to the inputs of others, the weight of each such connection can be positive or negative.

Connections with a positive weight are usually called exciting, with a negative one - inhibitory. Connections determine the computation of a neural network, and hence its memory and behavior. The principle is approximately the same as in a computer processor.

Back in the second half of the last century, it was proved that such neural network models have properties similar to the human brain. They can recognize patterns or, as mathematicians say, solve classification problems, but for this the neural network needs to be trained.

How pattern recognition works

For example, it is necessary that out of all objects on the Internet, the neural network is able to recognize one specific object. She is shown this object, in other words, information is fed into the input in the form of a given image. After calculating the values at the outputs of all neurons, the network gives a correct or incorrect answer.

If the network is wrong, the so-called error backpropagation algorithm calculates the contribution of each connection between neurons to the final error. Then, individually, he corrects their significance, after which the image is shown again by the neural network until it begins to accurately determine the given object in the picture, that is, how a person learns by repeating the material.

How long does a neural network take to learn?

No and yes. It all depends on the required accuracy

Although a young child is more intelligent than this high-tech device. In order to train the network, it needs to define several hundred thousand of these and other images.

At the beginning of training, the parameters of each element of the network, their weights, are set in an arbitrary way, which is why the network makes mistakes. In order to properly train the network and minimize inaccuracies, it is necessary to select such values of the weights using special algorithms so that the network works properly.

So, in order to train the network, it is necessary to select an example from the existing sample, show it to the network, get an answer and analyze the error. If the error is insignificant, the network can be considered trained, if the error is unacceptable, the weights should be adjusted and the training process repeated, again showing the network some examples from the training set.

To train artificial neural networks, programmers have come up with many complex architectural neural networks, the most promising of them are defined by the concept of DeepLearning.

This is the so-called deep learning process of multilayer neural networks. However, despite the deepest learning, it is still impossible to say that such artificial intelligence really thinks and makes decisions. This means that there is no need to fear that one day robots will decide to enslave or destroy humanity.

But it is already clear that in the future neural networks can serve a person well. They will help in medicine and safety, automate many different processes, become assistants in the house, replace consultants and managers, help to purchase a plane ticket or get the necessary advice.

The Topic of Article: How artificial neural networks work.
Author: Jake Pinkman