Deepfakes, hackers and COVID-19 diagnostics: what was discussed at OpenTalks AI

In early February, the OpenTalks AI conference dedicated to artificial intelligence was held in Moscow. We tell you what interesting things are happening in the industry and what are the main trends in our country and the world

Trends in the field of artificial intelligence

Igor Pivovarov, organizer of OpenTalks.AI and chief analyst at the MIPT Center for Artificial Intelligence, noted four main trends for AI and business:

Sergey Lukashkin, Director of Digital Transformation Project Management at VTB, spoke about how AI has affected business in 2020.

Here are the main factors:

• The volume of downloaded and stored data in the world has increased by 33 times: from 1,2 Zbytes (1,2 trillion GB) in 2010 to 40 Zbytes in 2020. These data are used as datasets for AI training.
• The growth of computing power – due to the fact that the cost of computing fell by 98%: from $1,8 per gigaflops (a unit of power, expressed in operations per second) in 2011 to $0,03 in 2020.
• The trend towards digitalization of business in all industries;
• Pandemic COVID-19.

What are the fastest growing areas of AI technology?

⚕️ Medicine and biology

Thanks to the pandemic and AI-powered technologies, this area has experienced an unprecedented breakthrough. For example, DeepMind’s AlphaFold AI has deciphered the process of protein folding, a task that scientists have struggled with for 50 years. With a 3D model of the protein structure, it is possible to treat many diseases and develop new drugs.

Artificial intelligence helps fight COVID-19. For example, the Open AI Consortium platform uses AI that learns from coronavirus medical data to predict the spread of the pandemic and the workload of medical staff, as well as suggest the most effective treatments.

Computer vision at the heart of the Care Mentor AI service platform, helps to detect signs of COVID-19 on CT scans with an accuracy of up to 90%, while doctors only up to 70%.

Alexey Chernyavsky, Senior Research Fellow at Philips Innovation Labs, said that neural network algorithms also help to get rid of “noise” and other distortions in CT images that occur due to movement, interference or metal implants. Computer vision removes artifacts and draws in missing data. As a result, it is possible to reduce the patient’s stay in the tomograph and reduce the radiation dose to 25% of the usual one.

Scanderm is a Russian application for online diagnostics of skin diseases based on a photo based on machine learning. The application, among other things, recognizes a rash with COVID-19.

Now AI is useful not only in the fight against the pandemic, but also with aging and complex diseases. For example, the GeroSense AI service helps fight aging. This is a smartphone application that collects indicators from fitness trackers – the number of steps, physical activity, pressure, pulse – and analyzes them over a certain period of time. Based on this data, GeroSence draws conclusions about the user’s biological age and develops personal recommendations for lifestyle and physical activity.

💣 Cyber ​​security

During the pandemic, the number of cyberattacks has increased markedly, but AI is helping to cope with them.

In particular, Andrey Arefiev, director of innovative products at InfoWatch, which specializes in information security in the corporate sector, said that machine learning helps to close vulnerabilities that are given very little attention in IT services. For example, some employees copy entire customer bases and technical developments before leaving – information security services rarely think about it. The damage from each such case can amount to the annual salary of the dismissed person. The algorithm recognizes typical signs that an employee is planning to quit: a decrease in productivity, visits to certain sites, and others. To do this, special software on the local network collects screenshots of the screen, monitors actions with databases, correspondence in instant messengers, and browser history. A month before the expected dismissal, AI notifies the IT service of possible risks.

👁️ Computer Vision

In 2020, facial recognition began to be used to combat the pandemic: to track patients in self-isolation, to identify people without masks or with signs of a cold. This spurred the development of technology.

DFI technology based on CV (computer vision) allows face recognition even when wearing masks. She does this with 14 points on her face that can be used to identify a person. Now the boundaries of privacy have narrowed even more.

Another promising area is 3D scenes, which are created using computer vision. Using neural networks, you can generate 3D models based on several photos or a panoramic shot.

The creation of 3D scenes is in demand in construction, interior design, military affairs, animation, as well as AR and VR. The NeRF neural network from Google Brain and Google Research recreates three-dimensional objects in photographs, completing their depth, color tones and texture, taking into account the point of view.

Hollywood uses technology to simulate lighting, actors and sets, saving on technically complex filming.

Read more in the material “What is computer vision and where it is used.”

Another trend is deepfakes – realistic videos that are created based on photos and videos of a real person, but put him in a fictional situation, sometimes even reproduce a fictional speech. Now they are becoming a new reality: the era of post-truth is coming, when generated images of famous personalities replace real ones. This poses new ethical problems for us. The Ministry of Internal Affairs of our country has already announced a tender for the development of technology that can recognize deepfakes.

🏭 Industry

Fully autonomous self-learning robots, systems based on the Internet of things, digital twins have appeared in production. For example, BMW and Nissan use AI-powered services during the final assembly phase to ensure that the car is free of defects and meets the required specifications. General Motors uses smart cameras for robots to detect defects and system failures at the assembly stage.

Another trend that is popular in the industry is generative design, when AI creates a product without human intervention. This allows the production of more durable materials, saving raw materials. Thus, enterprises become more autonomous, better adapt to market changes, predict quality and profitability. For example, Volkswagen uses generative design to create some car models.

🛒 Retail

Trade is increasingly moving to an online format, and in retail there is a steady trend towards contactless technologies. At the same time, online and offline are increasingly mixed due to the active development of delivery and applications that help make purchases in a regular supermarket. Thus, “smart” scales, price tags, chatbots and digital assistants, payment by face and augmented reality appeared.

The global trend towards stores without staff has reached our country: the first outlets have already launched Azbuka Vkusa and X5 Retail Group. Amazon Go has been operating in the United States for five years now – stores where you can select and pay for goods through an application that recognizes customers by their faces. The stores themselves also have cameras that identify visitors and monitor what goods they took and in what quantity. So far, the main problem of Russian projects is theft, but AI-based developments with face recognition help here too.

Sergey Zakharov, CEO of CERA, a developer of AI solutions for retail, spoke about how their face and object recognition system works in offline stores:

• monitors queues and redirects customers to another checkout to speed up the process;
• recognizes the products on the shelf and makes sure that they are correctly arranged (including the position of the label) and do not run out;
• observes customers, analyzing how they choose goods with reference to gender and age;
• helps track shoplifting.

📚 Education

The pandemic has led to the growth of online learning and the introduction of the latest technologies in the educational sphere. As a result, large amounts of data have accumulated on the Web – the so-called “digital footprint”. According to Sergey Lukashkin from VTB, there are AI-based algorithms that collect data about students into digital profiles. Then the technology analyzes them in several ways, comparing career successes and achievements. In the near future, this data can be used for personal recommendations: what a person lacks for a successful career and what are his prospects in the labor market. And also – to offer users online courses that suit them.

At the same time, the demand for professions that combine IT knowledge and soft skills, such as teamwork, analytical thinking, crisis management, as well as for professions that are related to AI and robots, has grown.

The main problems of online learning are the lack of live contact with the teacher, as well as control over how the student completes assignments and online tests. Maria Mashkeeva, Business Development Director at Electronic Platforms, presented Examus, a service based on face recognition, at the conference.

The neural network at the heart of the service observes students during an online exam: compares their face with a photograph and identifies them by handwriting to make sure that this is the same person. Examus notices if a student has left the frame, peeps or addresses someone else. The system reports all violations to the employee who conducts the exam and monitors compliance with all rules.

💳 Fintech

The main trends are seamless transactions and “invisible” banking: when you don’t even notice that you are using banking services. For example, Tinkoff Bank uses hyper-personalization: AI predicts the needs of the client based on data about him and offers the necessary services.

AI is also helping banks predict risk and analyze geodata to use to open new outlets and launch new products. In Sberbank, AI makes it possible to automate the process of issuing a loan by 95%.

⚖️ Jurisprudence

Unlike Legal Tech, which combines any online tools to automate routine actions, AI — that is, Legal AI — helps lawyers solve more complex problems and in 10-15 years may completely force them out of the profession.

However, for legal practice, most NLP solutions (NLP, Natural Language Processing – a range of tasks for processing texts in natural language) do not work due to language features and specific tasks. There are still no ready-made datasets for training neural networks in this area. Therefore, at the moment, the main task in Legal AI is to teach AI to understand the meaning of legal language.

NLP: what interesting happened in a year in text recognition

Grigory Sapunov, CTO of Intento, a company that develops AI solutions for working with content and translations, shared the main events in the field of NLP for 2020.

🧠 GPT-3 training breakthrough – a neurolinguistic network that generates coherent answers in a dialogue with a person. The amount of data and parameters used by it is 100 times greater than the previous generation – GPT-2. Now such neural networks are used, for example, by Sberbank.

👾 Trend for big transformers – that is, neural networks that use the attention mechanism, isolating important semantic pieces in the text. Also transformers are able to process different pieces of text. As a result, you can translate large volumes much faster, without waiting for the algorithm to read the phrase from beginning to end. Transformers are used, in particular, in Google Translate. Now they are successfully used in Yandex.Translate and Yandex.News.

📄 Improving popular chatbots. Google’s Meena or Facebook’s BlenderBot have improved significantly: they have learned to understand much more words and even take into account the context of the conversation. The Emora chatbot, 2020 Alexa Prize winner, set a new record: 20 minutes of meaningful conversation and a score of 4 on a five-point scale from the judges.

🌐 The global problems associated with AI have escalated. Creating and training neural networks costs millions of dollars, and so far only large corporations can afford it. Studies are also emerging that point to an environmental threat: CO emissions₂ in the process of training complex neural networks, six times the emissions from a car over its entire service life.

Subscribe also to the Trends Telegram channel and stay up to date with current trends and forecasts about the future of technology, economics, education and innovation.