How Deep Learning Works and Applications in the AI World

Among the various branches that make up artificial intelligence, there’s one that’s playing an increasingly important role: deep learning. Translated as deep learning, it’s the discipline that teaches computers how to process data, drawing inspiration from the structure of the human brain.

Many say it’s the true engine behind the most advanced technologies we use every day on our devices, such as facial recognition or automatic translations.

But what exactly is it? And how does it work? In this guide, we’ll look at everything you need to know about deep learning.

What is deep learning

Deep learning is an evolution of machine learning, meaning the ability of a machine to learn without being explicitly programmed. Deep learning is the method that enables this by leveraging hierarchical architectures that emulate the way the human brain behaves.

In technical terms, we’re talking about a subcategory of artificial intelligence based on Artificial Neural Networks with more than three layers. The term "deep" refers to the number of hidden layers where data is transformed.

This technology is actually the result of decades of research. As early as the 1940s and 1950s, researchers such as McCulloch and Pitts first, and then Frank Rosenblatt, designed similar systems such as the Perceptron, the ancestor of today’s networks.

In the 1980s, researcher Geoffrey Hinton popularized error backpropagation, a mathematical method for training multipath networks.

The turning point clearly came with the advent of AI, which added two key ingredients to all these notions: computing power and big data.

Deep learning was born to solve the problem of feature engineering. In older models, humans had to explain to the computer how to behave. Deep learning, however, is the designated intermediary, allowing the network to learn on its own which features are relevant by looking at millions of examples.

When we talk about deep learning

Deep learning begins to be discussed when three conditions related to the system’s complexity, architecture, and autonomy are met:

• Depth: if a network has multiple intermediate layers, from tens to thousands, the concept of machines creating increasingly abstract representations of data comes into play;
• No human intervention: the fundamental difference with machine learning is the absence of human input. Deep learning comes into play when the machine receives the raw data and independently decides which data is important for arriving at the solution;
• Big data: we begin to enter the realm of deep learning when traditional static methods no longer work. For example, if there are images and videos with unordered pixels, complex sound waves, or natural human language.

In practical terms, however, deep learning comes into play when the workload is so heavy that it requires dedicated graphics processors and chips. Deep learning, in effect, works every time a computer does something humanly complex, like understanding speech, driving a car, or generating an image.

How Deep Learning Works

To better understand how deep learning works, you can imagine an assembly line made up of small switches organized into layers.

First, there’s the layer decomposition. The initial layers recognize tiny, coarse details like points, lines, and angles. The intermediate layers combine lines to recognize shapes like circles and squares. The deep layers combine shapes to recognize objects like an eye, a wheel, or a nose. Finally, the output layers give the final verdict.

The second key step is weights, which are numerical values that connect the neurons to each other. If a neuron in the previous layer is important to the final decision, its weight will be high. Initially, these weights are random since the network doesn’t know anything yet.

The third fundamental step is learning. In the forward phase, the data enters, passes through the layers, is multiplied, and the network returns an answer. Then there’s the loss and output function, which measures the severity of the error. This way, the algorithm goes back and tells each neuron what’s wrong.

Finally, optimization. Using a method called Gradient Descent, the network mathematically searches for weight combinations that minimize the error. Essentially, deep learning works by trial and error. The more data it sees and the more corrections it receives, the more it improves.

What are the applications of deep learning

Today, the applications of deep learning are practically infinite. A prime example is computer vision, the ability of machines to see and understand the content of images and videos.

The same goes for natural language processing, which enables machines to understand, translate, and generate human language.

Its use in entertainment and media is increasingly widespread, for example when providing personalized recommendations on video or music streaming platforms.

Finance and security have also been using deep learning for years, to detect fraud by analyzing millions of transactions per second, and for algorithmic trading by analyzing historical market data to predict market changes.

Finally, there’s science and research, at the forefront with AlphaFold. This is a Google DeepMind system that solved the mystery of protein shapes, thus accelerating the search for new drugs. Even simple weather forecasts use models that process complex satellite data to predict climate events.

Some use cases of deep learning

To help you better understand what deep learning is and how it works, let’s look at some concrete examples of how this technology is used today.

We mentioned AlphaFold, a tool that solved a biological puzzle that had been under study for over 50 years. Scientists were trying to understand how proteins fold in on themselves, thus speeding up the creation of new drugs.

With deep learning, the system predicted the structure of almost all 200 million proteins known to science in just a few months. It has accelerated research into diseases like Alzheimer’s and the creation of new vaccines.

A second useful example is autonomous driving. To function, it must process x Gigabytes of data per second from cameras and sensors. With deep neural networks, training occurs using millions of hours of real-world driving. The car doesn’t follow rigid rules, but learns to recognize patterns, such as the difference between a shadow on the asphalt and a pothole.

A third example is real-time translation systems. Older translators simply replaced words in the selected language, but deep learning uses Transformers to analyze the entire sentence and context before translating. This way, the machine understands nuances and idioms to make the translation more fluid and natural.

There are dozens of other examples you use every day. Deep learning is used, for example, for facial recognition with FaceID, for airport security systems, for voice assistants like Siri and Alexa, for personalized recommendations on Netflix and YouTube, and for useful image creation tools like DALL-E and Midjourney.

Original article published on Money.it Italy. Original title: Come funziona il deep learning e applicazioni nel mondo AI