Through the development of modern science and technology, deep-learning technology, which is a field of artificial intelligence, has emerged. With the advent of this technology, the ability to educate computers to perform human-like tasks such as voice recognition and image classification has developed. In addition, a technology called “deep-fake” has been derived that synthesizes a person’s face or certain parts to create fake videos. This has a beneficial function of restoring people who cannot be met in the real world, but in another way, it has often been abused for crimes. In this regard, Pharos will examine the correct direction of deep-learning technology in the future.

 “Deep-learning technology,” simply put, is technology that allows computers to think and learn like humans. In other words, the computer has another consciousness. In detail, ‘Deep-Learning Technology’ is technology used to remember and classify objects or data by combining similar or related items. For example, on a computer, it is impossible to distinguish between an elephant and a puppy just by pictures. Deepꠓlearning technology enters a lot of data into a computer and classifies it among similar things to make distinctions. In other words, if a photo similar to a saved elephant and a puppy is entered, it is classified as an elephant or a puppy. The key to deep-learning technology is prediction through classification. Similar types of data are found in a large number of data, allowing computers to share data as if they were distinguishing between humans and objects. There are two ways to distinguish this technology: ‘supervised learning’ and ‘unsupervised learning’. The ‘supervised learning’ method is a method of teaching by entering information into a computer first. For example, being given pictures of elephants and recognizing them as elephants, the computer can distinguish pictures through pictures of elephants learned in advance. ‘Unsupervised learning’ simply means that there is no teaching process. It is a type of technology that allows a computer to learn photos by itself without prior learning. This is technology that is advanced compared to ‘supervised learning’ and requires a high computational capability. In particular, Google, which we all know well, has developed deep-learning technology to identify cat videos among videos registered on YouTube through the current ‘unsupervised learning’ method. In addition, deep-learning technology is introduced in the development of artificial intelligence systems of robots including voice recognition and translation, and deep-learning technology is used in SNS which we often use to apply to news feeds and image recognition fields.

 ‘Deep-learning technology’ is something that most people think is deep. However the word deep does not mean that you can get any deep insight. It shows the concept of learning what is stacked and expressed in a continuous layer. Recently, deep-learning models have dozens, hundreds, and tens of thousands of continuous layers for expression learning. These layers all apply to training data to automatically learn to help us reduce maximum error when we search and find the best picture or video for the commands we search for. In other words, ‘deep learning techniques’ are used to learn, stack, and process data in a number of stages.

The Advantages of Deep-Learning Technology

 The first advantage of “Deep-learning technology” is that it is possible for computers to learn faster than humans. It takes as much time and cost to develop experts in one field, but creating professional machines through deep-learning technology takes less time than cultivating talent, so experts in one field can be produced faster. For example, it is easy to understand if you think about the game between Lee Se-dol, a level 9 Go player, and AlphaGo, which was created through artificial intelligence technology on Google. The greatness of deep-learning technology can be seen by the fact that AlphaGo won 4-1 in the match against Lee Se-dol, who boasts the world’s best career, and Alpha Go, which has been in the world for less than a year.

 Second, customization is easier than with conventional rule-based algorithms. The existing rule-based algorithm requires a lot of experiments, manpower, cost, and time to be applied to the automation of the product, and the verification procedure that was preceded to verify the completed rules is required. On the other hand, ‘deep-learning technology’ allows people to quickly find the optimal rules on their own and find the most suitable content for the results of the search, reducing the number of things that people have to think about. Recently, ‘deep-fake technology’ derived from ‘deep-learning technology’ has become a hot topic. Simply put, ‘deep-fake technology’ refers to artificial intelligence learning data to produce fake products. The advantage of this technology is that it uses existing data, so it can produce videos or works without any new filming. It contributes greatly to the media entertainment industry due to its technological features that can reduce the cost of production and implement images without real-world photography. Furthermore, it can contribute to medical imaging technology based on X-rays such as CT and MRI, which can help to diagnose diseases and provide medical care. Also, people or objects that cannot be met in the real world can be restored using artificial intelligence technology so that they can be met through video. For example, late singers such as Elvis Presley and Turtleman, as well as the music of existing singers, can be expressed with ‘deep-fake’ music. Through this technology, there is a positive effect that can restore people who cannot be met in the real world and help us remember them.

 Finally, using ‘deep-learning technology’ has the advantage of reducing maintenance costs by saving power. For example, Google has a supercomputer on one floor of a building to process enormous amounts of data, and it takes a lot of maintenance to cool the heat generated by computers. However, the recent application of this ‘deep-learning technology’ has reduced the maintenance cost of cooling by 50% by processing only the data needed. In other words, ‘deep-learning technology’ has faster learning functions than humans, and can be easily and quickly customized. It has also led to advances in the media and medical technology industries, and has many advantages such as reducing technology and maintenance costs to restore people from the past.

The Disadvantages of Deep-Learning Technology

 We have looked at the advantages of deep-learning technology of computer artificial intelligence. Computers can classify data and store large amounts of data as easily as human brains, making it possible to use it in many ways. Many companies are trying to utilize deep-learning technology with a promising future. However, from a different perspective, some look at artificial intelligence technology negatively. ‘Moore’s Law’ states that the amount of data that can be stored on microchips doubles every 24 months. This leads to the rapid development of computing power. By allowing computers to have application power and find solutions on their own, they are worried that human beings will not need to order computers one by one, which could place a difficult problem on human abilities. Human ability changes slowly in the biological timetable. Therefore, the emergence of artificial intelligence that surpasses humans can be dangerous. Should a person’s attitude change in the development of technology, or should technology be controlled at the speed of a person? Experts argue that it is necessary to reflect on a unique human area that machines can’t invade.

 Deep-learning technology has several problems in the operation process. In the process of collecting-analysisꠓprocessing data, large amounts of data are required, creating difficulties. Difficulties also arise when applying the computed knowledge to other domains. First, the process of delivering the information received to the brain and then analyzing it is called ‘inference’. While people feel stimulated through the five senses, artificial intelligence lacks common sense skills, so shortcomings are clearly revealed when expressing reasoning or knowledge. The second occurs when a situation requires a large amount of data to be processed or when the situation acts on a complex area. Deep-learning techniques are necessary for simple answers to be applied or for immediate responses. Conversely, it is inappropriate if there is less data or if high-dimensional thinking is required. Deep-learning technology can still make hasty judgments in areas where errors have not yet been applied or where the results of the cause require a clear interpretation.

 For a while, a video of former U.S. President Barack Obama sharply criticizing Donald Trump was popular on YouTube. Barack Obama’s outspoken appearance was shocking enough to provoke a political controversy. However, the video was a manipulation video made with deep-fake technology. Deep fake technology uses deep-learning technology to create fake products by learning data on its own. Deep-fake technology makes it hard to tell what is real or fake. As time goes by, data that is difficult to distinguish is generated through repetition of the process of producing, identifying, checking errors, and determining results. Like former President Obama’s manipulated video, deep-fake technology is feared to be politically abused. More serious than this is the technology that is closely related to digital sex crimes. Recently, a terrible crime has emerged that uses deep-fake technology to synthesize and distribute the faces of acquaintances into pornographic materials in the ‘n-room incident,’ which has become a social controversy. There are no proper regulations on technology yet, so the severity of the damage will be greater. Currently, most of the deep-fake videos are increasing in the form of pornography, targeting female celebrities. The lack of recognition of digital sex crimes, which are considered okay because they are fake and not real, is also a problem. In the case of digital sex crimes, it is difficult to specify the perpetrator, and many unspecified people directly or indirectly commit the crime. Therefore, the vague fear and damage that the victim feels are so serious that it is impossible to recover. In order to solve the problem of technology abuse, developers are focusing on developing artificial intelligence that determines the authenticity of images around the world and are engaged in a fierce war with deep-fake technology.

An interview with a Kyonggi University professor (to get an expert opinion on deep-learning technology)

Q. Please introduce yourself.

A. My name is Lim Hyun-Ki and I am working as an assistant professor in the AI computer engineering department at Kyonggi University. I majored in artificial intelligence, received a doctorate degree and worked on artificial intelligence-related tasks at the Korea Institute of Science and Technology (KIST).

Q. Please explain about deep-learning technology.

A. Deep-learning technology is technology that belongs to artificial intelligence and machine learning, and the learning of artificial neural networks and the use of those neural networks are called deep-learning. In more detail, machine learning is divided into different categories based on purpose, technology, and among them, artificial neural networks are built by building multiple hidden layers of models for learning. An artificial neural network of tens or hundreds of hidden layers is called a deep neural network. Learning and utilizing this neural network is commonly called deep-learning. In the past, it was difficult to learn and use only a few hidden layers, but in recent decades, dozens or hundreds of hidden layers of neural networks have become available and used in many areas with high accuracy.

Q. What do you think of the abuse cases regarding deep-fakes using deep-learning technology?

A. As a researcher, it is regrettable that the finished technology is abused due to the development of technology. However, since any technology can be exploited, deep-fake can be considered as one of them. Perhaps the development of technology too quickly, especially in open-source software, has led to cases of abuse. With technology development, related laws, punishment regulations, and so forth should be quickly enacted, and social awareness should be raised regarding abuse cases and crimes associated with this technology.

Q. In what direction should deep-learning technology move forward?

A. I’d like to mention two of the many. The first is the social legislation we talked about earlier. Technology alone can create many problems by moving fast. We should try quickly to solve social problems that can occur with rapid technological development and not to cause them. At the same time as technological development, it will not be easy to contemplate these social problems and create solutions. However, I think that thinking about these things together can lead us to a faster self-driving car, a safer society, and so on.

The second is lessening the weight factor. In fact, Deep-learning technology is receiving a lot of attention and progressing in various directions, so it may be difficult to say that we should go in this direction. However I think lightweight should be the goal if you choose only one. One of the limitations of deep-learning technology is that it requires high computing power. This makes it difficult for ordinary people to actually access and utilize it. Reducing the size, complexity, or weight of artificial neural networks can actually increase usability. Of course, a lot of research is already being done to make it lighter.

 Artificial intelligence technology that creates virtual characters, including deep-fake technology that uses deep-learning, which is a key AI technology, has reached a level where we cannot distinguish between real and fake. Artificial intelligence technology was not designed to be used for crime at the beginning. It is the driving force behind the evolution of facial recognition technology and can be used as a positive means by improving facial recognition through deep-learning technology. Deep-learning technology’s ability to identify signs of cancer and symptoms of illness can be of great help to the medical field. The potential of deepꠓlearning technology for humans is huge, but if the problems that arise from it are also not recognized, they develop in an evil way. Due to the nature of deep-fake technology, where both positive and negative aspects coexist, it can be both the best and worst depending on the utilization. Discussions on the proper use of the double-edged sword of deep-fake technology and regulations need to be more active.

Reporter•KANG ROK KI•krk1754@naver.com

75th Reporter•KIM YUN JUNG•dbswnd7676@naver.com