In 2016, something happened that many experts considered impossible. AlphaGo, a program developed by DeepMind, defeated world champion Lee Sedol in a game of Go, a board game so complex that for decades it had been believed that no computer would ever be capable of mastering it. The surprise was not just the victory itself. Some of the moves made by the machine left even the world’s greatest players bewildered, who described them as creative and unexpected.

That day made it clear that artificial intelligence had entered a new era. Behind that demonstration was a technology that now powers ChatGPT, voice assistants, automatic translators and systems that help detect diseases: deep learning. But what exactly is it, and why is it transforming so many industries at once?

When computers only followed instructions

For much of the history of computing, computers could only do what programmers explicitly told them to do. If we wanted a system to recognise a cat in a photograph, we first had to define the characteristics it should look for: the shape of the ears, the whiskers, the snout or the tail.

This way of programming works very well when the rules are clear, but it falls short when dealing with much more complex problems, such as understanding a conversation, interpreting an X-ray or distinguishing one person from millions of faces.

The first major change came with machine learning. Instead of writing every single instruction, researchers began feeding computers thousands of examples so that they could discover patterns on their own. Even then, however, humans still decided which information was important.

What makes deep learning different?

Deep learning is an evolution of machine learning that removes much of this human intervention. Instead of telling the system what it should look for, it learns by itself which patterns are relevant.

To achieve this, it uses artificial neural networks, structures very loosely inspired by the way the human brain works. These networks are made up of multiple layers of interconnected neurons. The first layers detect very simple elements, such as lines, colours or contrasts. The following layers combine this information to identify more complex shapes, while the final layers are capable of recognising an object, understanding a sentence or predicting a response.

The word deep refers precisely to this succession of layers. The deeper the model, the more complex the patterns it is able to learn.It is this ability to learn automatically what is important that has turned deep learning into the driving force behind modern artificial intelligence.

Learning also means making mistakes

When someone learns a language or plays a musical instrument, they improve through repetition, making mistakes and correcting them. Artificial neural networks work in a similar way.

During training, the model analyses an enormous amount of data, makes a prediction and checks whether it is correct. If it is wrong, it slightly adjusts millions of tiny mathematical parameters that determine how it behaves. This process is repeated millions of times until the error becomes smaller and smaller.

That is why today’s models can recognise voices, translate languages or generate text with remarkable accuracy. Not because they “think” like we do, but because they have learned to identify extraordinarily complex statistical patterns.

The revolution is also about hardware

If this technology has existed since the 1980s, why has it only taken off now?

The answer is not only in the algorithms. It is also in the hardware.

Training a neural network requires executing trillions of mathematical operations. Traditional processors (CPUs) were not designed to handle this type of workload efficiently. GPUs, on the other hand, originally created to render video game graphics, are capable of performing millions of operations simultaneously.

When researchers discovered that these graphics cards were ideal for training neural networks, the development of deep learning accelerated exponentially. Companies such as NVIDIA, which for years had been associated almost exclusively with the gaming industry, became key players in the artificial intelligence revolution.

Without this evolution in hardware, ChatGPT or Gemini would probably still be laboratory projects.

A technology that is already part of our lives

Deep learning is far more present in our daily lives than we might imagine. It unlocks our phones through facial recognition, filters spam emails, detects fraudulent banking transactions, translates texts in real time and helps radiologists identify lesions that might otherwise go unnoticed.

It is also the foundation of the large language models that have popularised generative artificial intelligence. These systems do not search the Internet every time they generate a response. Instead, they use the patterns learned during their training to predict the most likely answer.

The same principle is also being applied to the development of new medicines, scientific research, autonomous driving and even the prediction of weather events.

The future also raises questions

Deep learning has demonstrated extraordinary potential, but it is not without limitations. These models consume vast amounts of energy, require enormous volumes of data and often operate as a “black box”: they deliver highly accurate results, yet even their creators cannot always explain exactly how they reached a particular conclusion.

There is also a strategic issue to consider. Developing the most advanced models requires infrastructures worth billions of euros, concentrating this technological capability in the hands of a very small number of companies. In a world increasingly dependent on artificial intelligence, controlling data, chips and computing power could become just as important as controlling energy or raw materials.

A revolution that is only just beginning

Deep learning is not simply a new computing technique. It is the breakthrough that has enabled machines to move beyond merely following instructions and begin learning from experience. This seemingly subtle difference is transforming entire industries and redefining the relationship between people and technology.

It is still too early to know how far this revolution will go, but one thing seems clear: understanding what deep learning is is no longer just a matter of curiosity for engineers. It is a way of understanding one of the technologies that will shape the future of the economy, science and society.

Protecting savings with physical gold has been one of 11Onze’s greatest contributions to its community, and its range of products continues to expand. In today’s environment of market volatility, persistently high inflation and growing distrust in the banking system, gold is once again strengthening its role as a safe-haven asset. Discover Or Llavor at Preciosos 11Onze.