What is Deep Learning? How Does It Work?

How does deep learning work and what is it? In the course of the article, we will talk about Deep Learning and understand its key concepts.

Subject of Study

The concept of “deep learning” has existed for 20 years, but people began to actively talk about it quite recently. Within the framework of this publication, we will explain what provoked active conversations around Deep Learning, what it is, what are the differences from classical machine learning, and other useful nuances.

What is it

Deep learning refers to one of the types of machine learning, it uses a special model, created by analogy with the brain – based on neural communication.

For the first time, the phrase was introduced into terminology in the 80s, but it became widespread only since 2012 before there were no computers capable of providing sufficient computing power. Due to a lack of performance, the learning model was neglected.

It gained popularity after a series of publications from famous scientists, a number of articles from scientific journals, and the like. When the technology began to develop, large media became interested in this area, for the first time they started talking about Deep Learning in the media of The New York Times. It was based on the material (scientific work) of specialists from the University of Toronto: D. Hinton, And Satskever, A. Krizhevsky. The work was based on analytical data on ImageNet image recognition. According to their observations, a neural network created on the basis of deep learning had a significant superiority. The efficiency of the system has reached 85%. This was the beginning of the spread of the neural network and its constant victory in competitions.

What is Machine Learning?

It is one of the areas of use of artificial intelligence, which describes how to create and build algorithms. It is based on the own experience of the program, that is, a special algorithm is not laid down by the programmer. The person remains indifferent, the machine independently determines the optimal way to solve the problem based on the transmitted data.

For clarity, consider an example: it is necessary to ensure the recognition of a person in photographs, the developer is required to provide about 10,000 pictures with marked human traits, then the program in the future will be able to determine patterns and identify the outlines of the body.

For teaching, a teacher is not always needed, sometimes the machine finds answers to questions on its own without outside help. It is noticed that the best results come when using a teacher. With each data processing, the system gains experience and becomes more accurate.

How Deep Learning Works

Its main task is to recreate the abstract thinking that a person possesses, then the computer will be able to generalize the parameters. An example, a trained neural network with the help of a teacher, does not understand well the handwritten font, which differs from person to person. To improve the results, the machine will have to provide all the ways of writing, only then can you count on the correct understanding of the handwriting.

Deep Learning is actively used when interacting with artificially created multilayer networks. The set goal for a deep learning system achieves the set goal much easier.

Today there are 3 main terms that coexist today and have approximately the same meaning: Deep Learning, Machine Learning, and Artificial Intelligence. In fact, these are different concepts, which are derived parameters from other properties:

1. Artificial intelligence is the presence of a wide variety of action algorithms designed to imitate the human solution of the assigned tasks. An example is a program – a simple chess game.
2. Machine learning is a branch of AI use, here the application not only solves the assigned tasks but also writes down the features of the solution for itself to create its own experience that simplifies and clarifies subsequent actions. For example – a chess application studies the opponent’s behavior and takes it into account for further rebuilding tactics.
3. Deep learning is one of the machine learning methods based on neural networks. When playing chess, it is primarily a neural network that learns.

How Deep Learning Works

Let’s illustrate everything with a simple example: we show the robot photos of girls and boys. Initially, the neural network is only trained to recognize changes in brightness. On the second layer of the network, it is already possible to recognize circles and angles. To the third circle – images of a person, gender differences are not yet taken into account, various inscriptions. With each subsequent circle, the recognizable patterns become more complex. Due to the neural network, the machine independently generates a representation, identifies important visual images, and even independently arranges them depending on their importance. In the future, the program will begin to better understand the image.

What’s Already Developed?

The largest number of projects today are using deep learning to recognize images and identify audio recordings, although there are already the first programs for diagnosing diseases. Already today, neural networks are used in Google for the translation of text from pictures. Using Deep Learning, it is easier to determine the presence of letters in photographs and their outline, and then the program translates the resulting text.

Another interesting project is DeepFace, it also specializes in working with photos, it was developed to determine facial features. Already today the accuracy of the program reaches 97.25%, the same accuracy is observed in humans.

In 2016, the WaveNet project from Google Corporation was launched – it is a system for simulating the human voice. To achieve quality learning, the system was loaded with millions of minutes of voice calls with the Ok Google service. After the entire learning cycle, the machine independently made a proposal, placed the correct stress everywhere, a characteristic accent, without any inappropriate pauses.

The neural network is even capable of semantic segmentation of video and photos, that is, the system learns about the presence of a person in the image and ideally accurately determines the contour of the face. Today, the technology is actively used in self-driving cars, the task of which is to detect obstacles on the road, markings, signs, and other road conditions.

A neural network in the medical industry helps to distinguish diabetic retinopathy only by providing photographs of a person’s eyeballs. In the United States, the technology is already in service in clinics.

Why has deep learning started to spread recently?

Previously, it was extremely difficult to apply the technology, costly, and took too long to learn. It all came down to the lack of power of graphics adapters and the amount of RAM. Due to the widespread availability of powerful GPUs, there has been a real boom in this area. Now they are able to work faster, are cheaper, and have virtually no storage limits.

This is a breakthrough, is it going to change now?

It is impossible to unequivocally answer the question, the experts did not come to a consensus. One side notes that billions of dollars in investment from Facebook, Google, and other giants make sense and will lead to even greater technology development. Deep learning is poised to transform the entire world of technology, according to optimists. Andrew Eung’s statement says “If the human mind is able to find a solution to a problem in a few seconds, there is a high likelihood that the process will be optimized soon.” This developer calls Deep Learning “new electricity”, comparing it to a major breakthrough for humanity. Most likely, those companies that will not implement deep learning will soon feel far behind the competitors.

There are also skeptics who argue that deep learning is nothing more than a buzzword. They assure that neural networks are just one of the methods of machine learning and far from the best.