A computational unit that combines input data to produce output information.
Artificial neurons are computational units that make up artificial neural networks. They are very loosely inspired by how biological neurons function, but in reality they perform mathematical operations.
Artificial neural networks (ANNs) are one of the main families of machine learning methods.
What is an artificial neural network?
An artificial neural network is a mathematical model capable of learning complex relationships from data.
It consists of small computational units called artificial neurons, organised into successive layers. Data is processed by the different layers of artificial neurons before an output is generated.
How does an artificial neuron work?
An artificial neuron receives several input values, each associated with a weight, which indicates its importance in the calculation.
The artificial neuron performs a mathematical combination of the inputs, then applies an activation function. The result is passed on to the neurons in the next layer, according to the following formula:
output = activation function (Σ input × weight + bias)
The most important parameters are therefore the weights and biases. These are progressively adjusted by the network during the learning process.
How does the network “learn”?
The network learns by comparing its predictions with the actual data, then generalising from the examples it has been given.
Take the example of a neural network designed to identify the presence of one or more tumours in MRI scans. The training input would therefore consist of an image in which the presence or absence of tumours is known, allowing the network to be trained on it.
- Based on this image, the network generates a prediction.
- This prediction is compared with the expected result (whether or not tumours are present).
- The network calculates a loss function, which measures the error made in its prediction.
- The network makes small adjustments to its parameters to reduce this error.
- This process is repeated across a very large number of examples.
The algorithm used to determine how the parameters should be adjusted relies in particular on backpropagation and optimisation methods such as gradient descent.
Through repeated training, the network gradually learns representations that enable it to perform the required task.
Why are multiple layers needed?
Multiple layers are needed to move beyond simple neural networks towards deep learning.
In a network with several layers, the first layers may learn to detect relatively simple features, while subsequent layers combine this information to identify more complex structures.
Taking the example of our previous MRI scan, the successive layers involved in identifying a tumour could be represented as follows:
- Pixels
- Edges
- Shapes
- Structures
- Object or abnormality.
However, this organisation is only a simplified representation: in a real network, the features learned are not necessarily so easy to interpret.
What are the main types of neural networks?
There are several architectures, each suited to different types of data.
- Convolutional neural networks (CNNs) are particularly well suited to image analysis. They are notably used in medical imaging to analyse MRI scans, CT scans and microscopic images.
- Recurrent neural networks (RNNs) and their variants, such as LSTMs, were designed to process sequential data, including time series, biological signals and text. However, they have largely been superseded by Transformers in many applications.
- Transformers are now essential for language processing and are also used for images, biological data and multimodal models. Large language models such as ChatGPT are based on this architecture.
- Generative networks, particularly GANs and diffusion models, are capable of generating new data, including text, images, sounds and other types of content.