Deep learning is more and more developed in our era of technological development. There is more and more talk of artificial neural networks capable of performances far surpassing those of a human brain. Here is what we need to understand about the computer system of artificial neural networks.


neural-network-1300px (1)

Deep learning

Deep learning is not only a question of technology, but also of a global vision of the human being. One of the primary objectives of deep learning is to reproduce the capabilities of the human being and even to create an improved version of them.


By deep learning, we mean a mode of automatic learning managed by a network of artificial neurons. The network is composed of several layers with precise functions.


Deep learning therefore tends towards the simulation of the human brain by activating these neuronal layers that interact and provide progressive learning from large volumes of data. Like the neurons of the brain, artificial neurons are able to communicate to promote learning and assimilation of various pieces of information.

The artificial neural network: what exactly is it?

Artificial neural networks copy the human brain to promote learning. It is therefore a system based on the functioning of the human brain to adapt it to computers equipped with artificial intelligence functions.


Thanks to the artificial neural network, the computer is able to solve problems autonomously. The network also improves the computer's capabilities.

The origin of artificial neurons

The concept of artificial neural networks originated in 1943 and was invented by neurophysicist Warren McCullough and mathematician Walter Pitts. Both were researchers at the University of Chicago. Their theory was presented in the Journal Brain Theory. In this article, the two researchers explain that the basic unit of brain activation is the activation of neurons.


By developing this theory, the researchers began to see the possibility of recreating the functioning of neurons. The year 1957 marked another milestone in the progression of artificial intelligence research with the invention of the perceptron. It is a binary classifier learning algorithm that represents the oldest Machine Learning algorithm. Based on this algorithm, machines will eventually recognize objects in images.


Today, the term perceptron also refers to a single-layer forward-propagating network.


Computers at the time were not powerful enough to process the amount of data needed to run the artificial neural network. However, the way to do this was known, and it was only a matter of time before researchers were able to get the learning process started. Due to a lack of resources, research into deep learning stagnated for several years.


With the years 2010 and the emergence of big data, the power to create sophisticated neural networks was obtained. It has even become possible to surpass the human being in the field of image recognition. The limits of neural networks are now constantly being pushed back, which makes it possible to envisage spectacular advances in the years to come.

robot-piano-1300px (2)

The functioning of the artificial neural network

The artificial neural network is based on several processors running in parallel. These processors are organized in tiers. The function of the first third is to receive raw data inputs. Each of the third parties then receives the outputs of information transmitted by the previous third party. The last third party is responsible for producing the results of the system. The more complex the problem, the more layers are needed to process it.

Each neuron has a particular value that determines what information can be transmitted to the system. The activation function calculates the output value of each neuron. It is this calculation that determines how many neurons need to be activated to solve the problem. An algorithm is then created. It matches a result to each of the inputs.

The algorithm allows the computer to learn from the new information it receives. The neural network in the computer makes the computer analyze examples so that it becomes capable of performing a task. These examples are labeled. It is this process that has made computers capable of recognizing objects in images, sometimes better than the human brain itself.

As with the human brain, artificial neural networks cannot be programmed directly, but must learn by studying and analyzing examples. There are three methods of learning:

  • supervised learning;
  • unsupervised learning;
  • and enhanced learning.

Supervised learning


For supervised learning, the algorithm trains using labeled data. To perform the task, it modifies itself until it can process the dataset to obtain the expected result. A concrete result must be defined for each input option. Supervised learning allows modifications to be made to the system in order to optimize the operation of the algorithm.


Unsupervised learning


In unsupervised learning, the neural network must analyze a set of data that is not labeled. A specific function tells it how far it is away from or close to the expected outcome. The network then adapts. The result of the task is therefore not determined in advance, but the system makes its own diagnosis based on the information obtained. The system is based, among other things, on the theory of adaptive resonance.


Enhanced learning


As for reinforced learning, it is a method by which one proceeds by reinforcements and sanctions depending on whether the results are positive or negative. Like the human brain, which learns by trial and error, the neural network learns progressively as it processes the data submitted to it.

The different types of neural networks

The different types of neural networks are usually defined according to the number of layers needed between data entry and the final result. The type of network is also determined from the number of hidden nodes in each model. The number of inputs and outputs of each node is also taken into account.

The basic neural network is called feed-forward. The information inside this type of network goes directly from the input to the processing nodes. From there, it is routed directly to the outputs.

As for recurrent neural networks, of a higher degree of complexity, they are able to save the results obtained after the information has passed through the processing nodes. The model is gradually fed and shaped from the saved results. Information can flow through a feedback loop and return to a previous layer. In this way, a memory is built up within the system.

As for convolutional neural networks, they allow the detection of simple patterns within an image to identify its content by cross-referencing. Their use is increasingly widespread in various fields, such as facial recognition and text scanning. They have a minimum of five layers. The result obtained passes from one layer to another.

Specialists also talk about forward-propagating neural networks. These networks transmit information in one direction only. Networks can have a single layer or several hidden layers.


To get the most out of your data, you should work with a professional like Ryax to help you get the most out of your data.

Do you do data analysis?
Discover our ready-to-use analyses on our marketplace.

La Ryax Team.