A big computer dream could soon see the light of day: computers with capacities far superior to those of the human brain. In recent years, research on artificial intelligence has made enormous progress. Artificial neural networks are a crucial technology for enabling machines to learn and “think” independently.
What is an artificial neural network?
Artificial neural networks constitute a specific branch of research in computer science and neuroinformatics. There are different types of artificial neural networks, offering different possibilities for processing information.
Artificial neural networks allow computers to solve problems on their own and enhance their abilities in general. Some require initial supervision, depending on the artificial intelligence method used.
How does an artificial neural network work?
The design of artificial neural networks is based on the structure of biological neurons in the human brain.
Artificial neural networks can be described as systems composed of at least two neurons – an input layer and an output layer – and generally comprising intermediate layers (“hidden layers”). The more complex the problem to be solved, the more layers the artificial neural network must-have. Each layer contains a large number of specialized artificial neurons.
Information processing within the neural network
Information processing always follows the same sequence within an artificial neural network: information is transmitted in the form of signals to neurons in the input layer, where it is processed. Each neuron is assigned a particular “weight,” and, therefore, different importance. Combined with the so-called transfer function, the weight determines what information can enter the system.
In the next step, a so-called activation function associated with a threshold value calculates and weighs the neuron’s output value. Depending on this value, a greater or lesser number of neurons are connected and activated.
This connection and this weighting draw an algorithm that matches the result of each input. Each new iteration makes it possible to adjust the weighting and, therefore, the algorithm so that the network gives each time a more precise and reliable result.
Artificial neural network: an application example
Artificial neural networks can be used in image recognition. Unlike the human brain, a computer cannot tell at a glance whether a photograph shows a human being, a plant, or an object. He is obliged to examine the image to discern its characteristics. The algorithm is put in place to know which factors are relevant; otherwise, he can find out for himself through data analysis.
Within each layer of the neural network, the system checks the input signals, i.e., images, broken down into individual criteria such as color, angles, or shapes. Each new test allows the computer to determine what the image shows more accurately.
Initially, the results are necessarily subject to many errors. If the neural network receives human origin feedback, which allows it to adapt its algorithm, we speak of machine learning. The concept of deep understanding aims to eliminate the need for human intervention. The system then learns from its own experience; it improves when an image is submitted to it.
In theory, we obtain on arrival an algorithm capable of identifying without error the content of a photograph, whether in color or in black and white, whatever the position of the subject or the angle at which it is viewed. Is represented.
The different types of artificial neural networks
Different structures of artificial neural networks are used depending on the learning method used and the objective sought.
Initially, the simplest form of artificial neural network consisted of a single neuron modified by weights and a threshold value. The term “Perceptron” now also refers to single-layer forward propagation networks.
Forward propagation neural networks
A forward propagating artificial neural network can only transmit information in one direction of processing. The systems can be a monolayer, that is to say, made up exclusively of input and output layers, or multilayer, that is to say, having a certain number of hidden layers.
Recurrent neural networks
It is possible to pass information through feedback loops in recurrent neural networks and thus bring it back to a previous layer. This feedbacks allow the system to build up memory. Recurrent neural networks are used, for example, in voice recognition, translation, and handwriting recognition.
Convolutional neural networks
A convolutional neural network is a type of multi-layered system. It is made up of a minimum of five layers. Pattern recognition is carried out on each of these layers. The result obtained on each layer is transmitted to the next layer. This type of artificial neural network is used in image recognition.
To ensure that the connections within an artificial neural network are correctly established, it is first necessary to “train” it. We can distinguish here two necessary procedures: supervised learning and unsupervised learning.
As part of supervised learning, a concrete result is defined for each of the different entry options. For example, suppose the system is presented with a photograph of cats for recognition. In that case, the system’s operation must be monitored, and feedback is given to determine whether the image is correctly recognized or not. This method makes it possible to modify the weights within the artificial neural network and optimize it.
In the case of unsupervised learning, the result of the task considered is not predetermined. The system itself draws the consequences from the only information entered, relying on Hebb’s learning rule or the theory of adaptive resonance.
Artificial neural networks and their applications
Artificial neural networks are a powerful tool in cases where we are faced with a large amount of data without knowing in advance where the solution should be directed. They are typically used in handwriting, image, and voice recognition, where a computer system searches for specific characteristics to assign them.
It is also possible to use artificial neural networks to carry out any prediction or simulation. This is the case, for example, for weather forecasts, medical diagnostics, or stock markets.
In industry, artificial neural networks are sometimes used as part of activity monitoring technologies to detect any deviations from determined values and automatically take the necessary countermeasures or independently set target values , taking into account the networks’ data assessment.
The development of unsupervised learning of artificial neural networks now makes it possible to expand their performance and their field of application considerably. The well-known example of Alexa, Siri, and Google’s voice assistant is the actual machine learning application artificial neural networks.
History and prospects
Artificial neural networks have appeared in the public sphere over the past decade as part of the artificial intelligence debate, but the technology’s creation dates back decades.
The first reflections on the subject date from the early 1940s. Warren McCulloch and Walter Pitts describe a model based on the structure of the human brain and connecting elementary units. The idea is then to be able to perform almost all the arithmetic functions. In 1949, Donald Hebb developed the learning rule that would bear his name, still used today in many neural networks.
In 1960, the first artificial neural network found worldwide commercial use by providing echo filtering in analog telephones. Research in this area subsequently came to a halt, linked on the one hand to the conclusions of eminent scientists who believed that the model of artificial neural networks could not be used to solve significant problems. On the other, This is because effective learning required large amounts of digital data, which was not the case. With the advent of big data, research will resume consistently, with interest in artificial intelligence and artificial neural networks, once again at their highest.
Since then, the field has continued to develop at a high rate. However, as promising as the results they provide, artificial neural networks are not the only technology to be implemented to support artificial intelligence in computing. They are only one option among many, although they are often presented in the public debate as to the only viable way forward in this area.