MMOOZZGG.com

Glossary

What is a neural network?

In short

A neural network is a program made of many simple calculation units, artificial neurons, that learns to find patterns from examples.

AI basics · neural network

MMOOZZGG.com editorsChecked: 1 min read

For a child

Remember how you learned to recognize letters. At first you mixed up similar ones, then you saw them many times and remembered them. Such a program learns in a similar way. Inside it has a great many tiny counters, and they pass signals to each other along a chain, like kids passing a ball around a circle. After many examples the counters get tuned, and the computer recognizes a cat in a photo or a letter in a notebook.

In plain words

A neural network is built like a long chain of filters. Numbers go in: the pixels of a picture, sound, or words turned into digits. Each artificial neuron multiplies the signals it gets by its own coefficients, adds them up and passes them on. At the output comes the answer: “there is a dog in the photo” or the next word of a phrase. At first the coefficients are random, and the answers are wrong. During training the network compares its result with the correct one and adjusts the coefficients little by little. Face unlock on a phone and word suggestions when you type a message work this way.

For business

Almost all modern AI services are built on neural networks: chatbots, document recognition, call analysis, image generation. Companies usually do not build their own from scratch: they take a ready one and connect it to their tasks through the cloud or install it on their own server. The cost depends on the model size, the volume of requests and the rules for storing information: if customer data cannot be sent abroad, you need an option in Russia or inside your own environment. The risk is that a network makes confident mistakes and does not explain why it answered the way it did. So quality is checked on the company’s real examples before launch, and important decisions stay with a person.

For a specialist

An artificial neural network (ANN) is a parametric function made of layers: each layer applies a linear transformation to the input with weights and a bias, then a nonlinear activation function. Feedforward networks (MLP), convolutional networks (CNN), recurrent networks (RNN) and transformers differ in the pattern of connections between neurons. Weights are fitted to data: a loss function is minimized by gradient methods, and gradients are computed by backpropagation. D. Rumelhart, G. Hinton and R. Williams showed its use for multilayer networks in Nature on 9 October 1986: hidden units learn by themselves to represent the important features of a task. Networks with many layers belong to deep learning; the 2015 review by Y. LeCun, Y. Bengio and G. Hinton in Nature describes how such layers build more and more abstract representations.

Example

A warehouse scans delivery notes, and a neural network finds the number, date and amount in them by itself. The operator checks only the disputed fields instead of retyping the whole document.

Related terms

Questions and answers

What is a neural network in simple words? It is a program that learns from examples, not from rules written in advance. It is shown many pairs of “question and correct answer”, and it adjusts its internal settings by itself to answer in a similar way.
How does a neural network differ from artificial intelligence? Artificial intelligence is the whole field, and a neural network is one of its tools. Today this tool is the most common, so in conversation the words are used as synonyms.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us