What are neural networks and how do they work?

In recent years, neural networks have gained considerable importance in artificial intelligence projects. As a matter of fact, companies like Google and universities like Stanford have been able to improve machine learning processes by using neural network algorithms.

Even though neural network theory is not a recent phenomenon, most people are still unfamiliar with it; in addition, it is frequently confused with other fundamental concepts of artificial intelligence, such as machine learning or deep learning. In this article, we explain what neural networks are, how they work, and the applications that they have already been used for.

What are neural networks?

Neuron networks, as their name implies, are a method of calculating that emulates the way neurons function in an organism. Using that inspiration he draws a model of interconnected units that generate, transmit, and reinforce concepts to reach certain conclusions and consolidate them as knowledge.

In artificial intelligence, neural networks are nothing more than mathematical representations of biological processes. The same way the brain learns patterns in grammar when learning a language, the computer processes specific data and finds the most appropriate parameters to process them.

Neural network models are used by computers to find suitable parameters and patterns through “training”. By feeding the computer with data, the computer determines valid analysis criteria and applies them. Neural networks, for example, can be used to identify bicycles in photographs, after identifying the geometric shapes that comprise them.

How do neural networks work?

So, how do neural networks work? As we have already mentioned, its structure tries to emulate the neural networks that make up our own brain. The neural network is made of neurons and perceptrons, each of which receives specific input data or values.

Upon receiving these input values, the perceptrons modify them according to their own weight, which is the influence they will have in determining the conclusion we want to draw. These modified values are then moved to the next neurons that make up the network. In this way, the values are modified until a conclusion is reached.

Differences between machine learning and neural networks

In summary, neural networks are an algorithmic model for pattern recognition, which allows a computer to learn. Since they are algorithmic, they are often confused with machine learning, which covers a broader range of applications.

Machine learning, on the other hand, encompasses all those techniques and methods that are available to teach computers to learn. Neural networks are just one of these methods, but there are several more: decision trees, probabilistic models, linear models and a long etcetera.

How to create neural networks?

It is easier to understand how to create neural networks by simplifying a real example. One of the most successful neural networks is the recommended videos section of YouTube. The web gets a lot of data that can be used to learn: the time we spend watching a video, the videos we choose, the ones included in our playlist, the videos we’ve “liked” … YouTube’s mission is to learn our tastes from these parameters.

As we have mentioned, the fundamental unit of a neural network is the perceptron: an element that connects with several inputs, each of them with a particular “weight”. In our example, let’s say that initially YouTube values all the parameters that we have mentioned as equally important. Obviously, clicking on a video is not as important as marking it with a “Like”, so after a while it will stop recommending videos that interest us. Depending on the results, the weight of the entries will have to be changed. This first phase of learning is known as feedback or adjustment.

Things become more complicated when we incorporate the concept of a multilayer neural network, which consists of adding more neural layers to achieve a better result. In our example, we add one more perceptron in charge of finding out how many recommended videos have been marked with a “Like”. If no video has been marked, the perceptron determines that these recommendations are not valid and that therefore it must modify the weight of one of the inputs.

Of course, the previous example is a simplification: perceptrons are only one of the elements that are used, there are much more complex neural network models, such as convolutional neural networks. These types of networks are often used to identify objects in images, through the search for specific characteristics in small groups of pixels. Instead of tracing the entire image, they pay attention to shapes of the same colour in a particular space in the image. Naturally, the more neural layers added to the network, the more details they can find and the easier it will be for them to learn to distinguish a particular image.

Uses of neural networks

We noted at the beginning of this article that neural networks have a long history: the first theories about them appeared in the second half of the 20th century. However, the most recent technological advances are what have allowed this machine learning model to obtain good results and begin to be applied.

Besides the YouTube recommendation algorithm, there are many other examples of the neural network application. Without going any further, Google has used this methodology for two different projects: the use of neural networks allowed it to pass the famous reCAPTCHA image identification test. In addition, the company has also used a convolutional network to recognize street numbers in Street View.

Academic institutions have also developed new machine learning techniques using neural networks. By automatically identifying images, Stanford was able to automatically generate captions. The University of Edinburgh used convolutional neural networks to detect successful strategies in the game Go.

Most previous projects show that neural networks are still under development and experimentation. Even so, this methodology continues to spread in popularity, suggesting it won’t be long before it is widely adopted. Everything indicates that neural networks will soon become one of the main pillars of artificial intelligence.