We are already accustomed to the phrase “artificial intelligence”, to what a smartphone or TV offers
But this does not mean that the capabilities of technologyexhausted or limited only by simple everyday life. The pandemic has caused a new leap in the development of AI and machine learning (ML) algorithms, also due to dramatically changed economic conditions and people's habits. Media consumption has skyrocketed, with social isolation helping. All this has required more personalized interactions with customers, media and entertainment (M&E). For example, Netflix, with the help of AI algorithms, was able to not only maintain the quality of streaming, but also significantly improve it: the platform predicts the future needs of viewers and places resources in strategically important server locations. By pre-positioning video assets closer to subscribers, users can stream high-quality video even during peak hours. But, of course, the possibilities of artificial intelligence are much wider and have virtually no boundaries: from advanced quantum computing systems and medical diagnostics to consumer electronics and smart personal assistants. If in doubt, ask Alice or Siri.
Experts believe that 80% of technologies thatwill be developed in the coming years, will be based on AI algorithms and ML. The number and variety of artificial intelligence applications continue to grow, and researchers and scientists are constantly finding new ways to use them. According to research, today 77% of the devices we use in life have built-in artificial intelligence.
AI is spreading at a fast paceboth at the chip level and at the software level. Moreover, both directions are closely related to each other. Manufacturers such as NVIDIA, Intel and Qualcomm are actively improving hardware, making AI systems faster and more complex. This leads to a greater democratization of AI. More and more software developers and corporate IT employees can use artificial intelligence algorithms when working with data. This is already noted by many cloud service providers: AWS, Azure, Google, Oracle and IBM. They are embedding and expanding their AI offerings for public and hybrid cloud deployments. Ultimately, this means greater availability of the computing power, frameworks, and algorithms needed to apply AI to everything from smart speaker to mobile device to enterprise sales and scheduling software.
How AI works with texts and why it helps older people
Generative models based onconversational AI, in the era of a pandemic have become more in demand than ever. The reason is simple - living and working remotely has caused people to request personalization when using digital resources. Until recently, only a person could create such an experience. Now it has been replaced by chatbots and natural language processing (NLP) technology.
Recent ResearchandMarkets Report showedthat the global conversational AI market is expected to grow from $ 4.8 billion in 2020 to $ 13.9 billion in 2025.
NLP technology allows you to imitate humantalk. And chatbots working on its basis are today one of the most popular ways of personalization and cost optimization: their implementation allows many enterprises to reduce costs by up to 90%. But the biggest value of conversational AI is that it enables personalized communication. It can be trained to be multilingual or even provide empathetic support to the user. Intelligent chatbots can, for example, help older people cope with loneliness. Osmar Zayane, an artificial intelligence expert at the University of Alberta, for example, led a project that aimed to develop a chatbot that could simulate dynamic conversation and provide social gratification for older adults experiencing loneliness.
One of the most effective language models isTransformer. Google is actively experimenting with this method of generating text. Earlier this year, the company announced that it was able to train a model containing 1.6 trillion parameters. In April 2021, the Google record was broken by the research group of the Chinese company Huawei, which announced the creation of the Chinese equivalent of GPT-3: the 750 GB model, called PanGu-Alpha, contains up to 200 billion parameters - 25 million more than GPT-3. and was trained on 1.1 TB of e-books, encyclopedias, news, social media and web pages.
The "rarely activated" method used inits models Google and Huawei, unites several models within a more global one, and also allows you to build in a strobe network, which decides which model to apply in each case.
During the experiment, the researchers askedtrained models the task of predicting words in passages. At the same time, about 15% of the words were missing in the text. However, Google does not deny that the use of AI to generate text is still not fully adapted to the real world. First of all, due to the presence of prejudices and various types of xenophobia, which cause stereotyped thinking in artificial intelligence. For example, the AI model can put the adjective “naughty” next to the word “woman” or tell the patient to kill himself, as it was during the experiment of the French company Nabla.
The GPT-3 model, introduced last year, has successfullycoped with what was previously considered an exclusively human foundation of work - she wrote an article for The Guardian, in which she told why AI does not threaten humanity, and also learned how to translate texts, answer questions, write poetry and prose.
Sber has applied a similar teaching methodologylanguage model for your assistant. The model was trained on Russian literature, supplementing its knowledge with a dataset with dialogues. In addition, Joey's assistant has a built-in ranking mechanism that allows him to choose the most interesting answers. Moreover, Joy does not choose pre-prepared replicas, he builds phrases in real time. Therefore, communicating with him looks like communicating with a person.
Where artificial intelligence algorithms are already being actively used
- cinema and TV;
- personalization of user experience;
- social media;
- journalism;
- music;
- games;
- sport;
- medicine;
- cybersecurity;
- fighting deepfakes;
- automation and personalization of production;
- collection and processing of information.
How AI helps humans create content
AI algorithms help people develop theirtalents, creativity. What machine models are always criticized for is the lack of ability to create what a person cannot do. But they easily allow a person to expand his scope for imagination. In the Internet space, users now think about the image they present to people, about the content. To gather as many subscribers as possible, you need a high-quality product, unlike anything else, and at the same time characteristic of the author. At PicsArt, we actively use AI so that users can work with images without any limitations. Algorithms help us make complex changes, such as changing the background, removing unnecessary objects, improving the quality of images and changing their style. This also allows us to improve the overall user experience.
All metadata we collect is usedto directly improve the user experience. It's a virtuous cycle: anonymized, privacy-compliant user data helps us improve our product, a better product increases usage, and more usage generates more data, making our AI even smarter. This cycle is essential for the massive growth of a business like ours.
Plus, AI helps PicsArt usersto simplify their work: for this, the service implements systems for searching content by tags, recommending stickers and searching for similar images, which selects photos by the most common colors or by the description of the plot in the images. There are models who simply rate photos for similarity.
If we talk directly about unusualways of working with images, then, of course, these are now popular processing - turning a photo into a cartoon or anime, applying effects and visual solutions, such as Canvas, Sketch effect, Style transfer, Upscale, or improving an image according to technical and artistic criteria. The goal, in any case, is the same - to create content that will attract more attention.
Users love that they can useeasy-to-use tools to make paintings from your photos that look like the work of great artists. Essentially, become a digital artist. But behind this lies the work of deep learning models.
To explain how suchmodel, an analogy can be made. Imagine a situation when you are given two pictures: your photo and a painting by an artist, and then asked to draw a photo, but with the help of paints and colors from the picture. How would you do it? For example, I would try to sketch with a pencil, and then try to color it in the artist's style, but without forgetting about the sketch itself.
One of the PicsArt art effects - Canvas - based onAI algorithms selects a famous painting or sculpture from the times of Antiquity, the Middle Ages or the Renaissance for a photo uploaded by the user. Using face recognition technology, the art effect creates a double image of a person and a work of art. To create the Canvas, over 6,000 tilt and face experiments were carried out to find the optimal combination of elements. To train the neural network, it took a dataset with over 2,000 pieces of art.
Artificial intelligence helps andprofessional photographers who have to process hundreds of photos. IT giant Adobe uses an artificial intelligence engine in its Sensei product. It is capable of analyzing photographs and comparing them against a database of thousands of professionally edited images. Based on this analysis, he intelligently recommends the most appropriate editing and adjustments for your shot.
Luminar AI photo editor also uses AI thatcan be seen directly from its name. True, some users consider the editor's approach too automated, but the editor's tools, according to the developers, will allow you to retouch faces without difficult and demanding operations, add weather conditions to the photo and adjust colors and lighting for them. The Composition AI model automatically aligns images and suggests cropping based on composition guidelines and feedback from professional photographers.
How AI determines if a photo or video is a deepfake
AI algorithms gave birth to deepfakes and now they themselvesthey are fighting. This area is one of the priorities in cybersecurity. Using the faces or votes of top managers is a new type of fraud. But unlike sophisticated technologies such as ransomware, deepfake attacks rely on social engineering: they rely on deception. According to ZDnet, the average loss per complaint for such attacks is $ 75,000. The average loss from malware per complaint is $ 4,400. This is why researchers at the Dawes Center for Future Crimes at University College London rated deepfakes with simulated human audio and video images as the most dangerous criminal threat associated with artificial intelligence.
History really plays tricks onAI developers. Stanford's Manish Agrawala two years ago helped develop lip-syncing technology that allowed video editors to almost seamlessly change the words of speakers. The tool could easily insert words that the person had never said, even in the middle of a sentence, or delete words that the person had said. To the naked eye, and even to many computer systems, everything looked organic.
But this technology has created tremendous opportunitiesfor scammers, political blackmail and crime. For example, in Russia, scammers have created a deepfake copy of the founder of Flocktory and Dbrain Dmitry Matskevich. For almost half an hour's video, deepfake-Matskevich talked about the platform with a new earning system. Of course, the domain linked to in this video belonged to cybercriminals.
Therefore, a year after the end of developmentAgrawal's lip sync technologies have introduced an AI algorithm that can detect deepfakes in video. The program accurately detects more than 80% of fakes, recognizing the smallest inconsistencies between the sounds made by people and the shape of their mouth.
But, according to Agrawal, long-termthere is no technical solution for finding and identifying deepfakes. Technologies for their creation also do not stand still: today, given a sufficient number of samples of a person's face and voice, the creator of a deepfake video can make a person “say” anything.
Agrawal's tool works on the basis ofAn AI algorithm that looks for inconsistencies between “visemes,” or mouth shapes, and “phonemes,” phonetic sounds. In particular, the researchers looked at a person's mouth when he made the sounds "B", "M" or "P", because it is almost impossible to make these sounds without closing the lips tightly.
AI algorithms will continue to actively develop,offering users of digital services more and more options: from ensuring safety and improving the quality of medicine to creativity and voice assistants. The introduction of AI will go on more and more actively, and the market will develop.
Last year, OpenAI gave the biggest leap innatural language processing. However, this model of artificial intelligence required an enormous amount of computational resources. Microsoft plans to help OpenAI work together to leverage the company's supercomputers to create even more powerful and reliable AI models. Most likely, more emphasis will be placed on AI, which will also help optimize and reduce the power consumption of these data-hungry machines.
Google DeepMind, AI for Good by Microsoft,Facebook AI, Intel University Research & Collaboration Office (URC), NVIDIA AI and OpenAI are just some of the most prominent companies and organizations that conduct AI research. They will help people solve many problems related to health, poverty, education, the environment and everything else that concerns our lives.
Read also:
Japanese scientists drilled the ocean floor near Fukushima at a depth of 8,000 meters
A mathematical model of the brain will allow AI to think like a human
Uranus has received the status of the strangest planet in the solar system. Why?