Last reviewed:
What is an artificial neural network? Simple definition and types (CNN, RNN, transformer)
An artificial neural network is a computing model made of small interconnected units, artificial neurons, organised in layers: each neuron combines its inputs according to weights, then passes a signal to the next layer. Learning consists of automatically adjusting these weights, from examples, until the network produces the right answers.
France's official definition (Journal officiel, December 9, 2018) describes a “set of interconnected artificial neurons that constitutes a computing architecture”. An artificial neuron is a device with several inputs and one output: it multiplies each input by a weight, adds everything up and applies a non-linear function, often a threshold. The model dates back to 1943: Warren McCulloch and Walter Pitts described a neuron that forms a weighted sum of binary signals. Frank Rosenblatt proposed the perceptron in 1957, the first network able to learn its weights, published in 1958. A network stacks layers: an input layer (the pixels of an image, the words of a text), hidden layers, an output layer (the answer). The weights are the model's parameters; a large language model has billions of them. Training adjusts these weights through backpropagation, a method demonstrated in 1986 by Rumelhart, Hinton and Williams: the error at the output is measured and each weight is corrected in the direction that reduces it. A network with many layers falls under deep learning. Three main families dominate. Convolutional networks (CNNs), developed by Yann LeCun in the late 1980s, analyse images. Recurrent networks (RNNs), including the 1997 LSTM, process sequences such as speech. The transformer, published in 2017, has largely replaced them for language. The brain image remains an analogy: an artificial neuron is only a mathematical operation.
Concrete example
Illustrative case: a French accounting firm with 60 staff processes 15,000 supplier invoices a month for its clients. Its automated data-entry solution chains two neural networks. A convolutional network reads the scanned image and locates the text areas. A transformer-type model interprets this text and extracts the supplier, date, net amount and VAT. Staff now only check invoices where the system flags a doubt. The firm never sees the model's weights: it evaluated the solution on a sample of its own invoices before adopting it, and measures the error rate every quarter.
Comparison
| Multilayer perceptron | Convolutional network (CNN) | Recurrent network (RNN, LSTM) | Transformer | |
|---|---|---|---|---|
| Origin | Perceptron (1957-1958), multilayer training (1986) | Yann LeCun, late 1980s | LSTM by Hochreiter and Schmidhuber (1997) | Google researchers, 2017 |
| Principle | Fully connected layers | Filters that scan the image region by region | Sequential reading with memory | Attention: each element weights all the others |
| Suitable data | Tabular data, small problems | Images, video | Speech, time series, text | Text, code, images, audio |
| Business use | Scoring, simple classification | Quality control, document reading | Forecasting, speech recognition (historically) | Generative AI assistants, translation, extraction |
FAQ
What is a neural network, simply put?
It is a program made of thousands or billions of small interconnected computing units, organised in layers. Each unit performs a simple operation: it weights what it receives and passes the result on. Together, and after training on examples, they can recognise an image or understand a text.
How does a neural network work?
Data enters through the first layer, passes through the hidden layers, where each neuron combines its inputs according to weights, and exits through the output layer. During training, the output is compared with the right answer and the weights are adjusted to reduce the error (backpropagation), millions of times.
What is a convolutional neural network (CNN)?
It is a network specialised in images. It slides small filters over the image to detect local patterns (edges, textures), then increasingly complex shapes. Developed by Yann LeCun in the late 1980s, CNNs are used for image recognition, quality control and document reading.
What is the difference between a CNN, an RNN and a transformer?
A CNN analyses images region by region. An RNN reads a sequence element by element while keeping a memory, which suits speech or time series. The transformer, published in 2017, processes the whole sequence at once thanks to the attention mechanism: it underpins today's language models.
What is the difference between a neural network and deep learning?
The neural network is the tool; deep learning is the approach that uses networks with many layers. A small network with a single hidden layer is a neural network, but it is not called deep learning.
Does a neural network work like the brain?
Only by analogy. The artificial neuron is inspired by the biological neuron, but it is just a mathematical operation: a weighted sum followed by a function. A neural network does not think or understand in the human sense; it detects statistical regularities in its data.
See also
Further reading
Sources
- Vocabulaire de l'intelligence artificielle, official French definitions of “neurone artificiel” and “réseau de neurones artificiels”, Journal officiel No. 0285, December 9, 2018. https://www.legifrance.gouv.fr/jorf/id/JORFTEXT000037783813
- Scientific Background to the Nobel Prize in Physics 2024 (McCulloch and Pitts 1943, Rosenblatt 1957, backpropagation 1986, CNNs, LSTM), Nobel Committee for Physics. https://www.nobelprize.org/uploads/2024/09/advanced-physicsprize2024.pdf
- Fathers of the Deep Learning Revolution Receive ACM A.M. Turing Award (backpropagation, convolutional networks), ACM. https://awards.acm.org/about/2018-turing
- Intelligence artificielle, de quoi parle-t-on ? (explainability of neural networks), CNIL. https://www.cnil.fr/fr/intelligence-artificielle/intelligence-artificielle-de-quoi-parle-t-on