От робототехники до здравоохранения и дизайна париков, технологии искусственного интеллекта (ИИ) оказали влияние на широкий спектр отраслей. Согласно анализу данных  компании Crunchbase, доля финансовых инвестиций, получаемых стартапами в области ИИ в последние годы колеблется от 9% до 10%. Самый большой объем финансирования приходится на робототехнику, автономные транспортные средства и корпоративное программное обеспечение. При этом инвестиции быстро увеличиваются также и в другие отрасли. ИИ находит все более широкое применение в биотехнологии, компьютерном зрении, лингвистике, финтехе, разработке полупроводников, безопасности, логистике, обучении, агротехе и т.д. Внедрение и использование технологий искусственного интеллекта — не только создают возможности развития нового бизнеса сами по себе, но и применение их для решения различных бизнес-задач, создания новых продуктов и услуг — представляют широкие возможности для предпринимательства.

Back in 2013, funding for AI companies totaled just $3 billion with fewer than 1,000 deals. In 2021, AI funding peaked at $69 billion and the number of deals quadrupled.

In this post, we will highlight some of the rapidly growing sectors in this area.

Synthetic data in AI models

Synthetic data is information that is created artificially rather than as a result of real events. Synthetic data is generated algorithmically and is used as a replacement for test sets of production or operational data, for testing mathematical models and, increasingly, for training machine learning models.

Benefits of using synthetic data include reducing the limitations of using sensitive or regulated data, tailoring data needs to specific conditions that cannot be achieved with authentic data, and creating datasets for software testing and quality assurance for DevOps teams.

Companies from healthcare to finance are experimenting with synthetic data sets. In industries where there is not enough real-world data to train AI, or where compliance and privacy are major concerns, enterprises are turning to synthetic datasets: artificially generated images, videos, or tabular data that are similar to real-world datasets.

Disadvantages of synthetic data include inconsistencies when trying to reproduce the complexity found in the original data set and the inability to directly replace authentic data, since accurate authentic data is still required to obtain useful synthetic examples of information. Despite this, many companies are actively using this technology. It is most widely used in medicine, finance and telecommunications.

Company Illumina uses synthetic genomics data developed by startup Gretel, for medical research. The two companies concluded that legal requirements and the need to obtain patient consent significantly limited the speed and scope of medical research based on real patient data. To get around this problem, Gretel used real genotype and phenotype data to train an AI algorithm that generated a set of artificial genomic data.

In the financial sector JP Morgan  trains AI financial models using synthetic data.

It is estimated that up to 85% of real-world customer data in the telecommunications industry is unusable due to a lack of customer consent, making behavior analysis and prediction difficult. So company Telefónica together with a startup Mostly AI generates synthetic client profiles that reflect statistical patterns of real data.

Who will win the race to create chips with integrated AI?

The rapid commercial success of AI across industries is creating growing demand for specialized hardware that can handle resource-intensive AI workflows. Such needs exist both in cloud data centers and in peripheral devices such as photo cameras.

Previously, this market was clearly dominated by NVIDIA GPUs. But with growing demand, new players appeared in the market.

Например, последние модели смартфонов Google Pixel оснащены собственными процессорами производителя — Tensor. Они предназначены для поддержки приложений искусственного интеллекта на устройстве. А Amazon в четвертом квартале 2021 года выпустил процессор Graviton3, свой собственный чип для логических операции ИИ в AWS.

Startups are also joining the race. Company Cerebras Systems предлагает «Wafer-Scale Engine (WSE) — революционный центральный процессор для компьютерной системы глубокого обучения. “…это самый большой из когда-либо созданных компьютерных чипов и самый быстрый процессор искусственного интеллекта на Земле.. В отличие от процессоров общего назначения, WSE был создан с нуля для ускорения глубокого обучения: 850 000 ядер для тензорных операций, огромная встроенная память с высокой пропускной способностью и соединения на несколько порядков быстрее, чем традиционный кластер” — говорится на их сайте.

However, giant microprocessors are not suitable for many everyday AI applications due to size limitations and power requirements. Accordingly, more and more companies are offering AI processors that can be used with peripheral devices such as automotive sensors, cameras, robots, etc.

Startups are working on this problem, including Mythic, Syntiant And Kneron, who managed to raise more than $100 million each to develop this technology.

Компании Untether AI и HOUMO.AI, работают над другим подходом к решению задачи — создание комбинированных микросхем, объединяющих процессор ИИ и память компьютера — «вычислениями в памяти». Высокий уровень интеграции в таких системах может обеспечить значительный прирост производительности по сравнению с традиционными системами. Компания Samsung reportsthat this approach allowed her to more than double the speed of the speech recognition neural network while simultaneously cutting its power consumption by half.

Another company Graphcore  also tries to improve performance by working with the processor structure. It uses an approach called "processor 3D". This technology involves connecting multiple sets of chips together to create an integrated stack. Among other things, Graphcore included a chip in the stack that manages power consumption. As a result, the neural network training process occurs at 40% is faster and with lower energy consumption.

Other companies are doing away with the physics of conventional AI chips entirely. They are developing chips based on photonic processors that transmit data using light rather than electrical signals. The big advantage of photonics is speed—light can transmit information faster, with greater bandwidth, and with less energy consumption than electrons. The demand for AI-powered data processing is expected to increase exponentially, and the use of photonic technologies will push the hardware performance limit.

Startups Lightmatter And Luminous Computing , are developing photonic chips optimized for tasks such as deep learning. Based on such an elemental base, they plan to create “supercomputers with artificial intelligence” capable of processing particularly complex algorithms that modern supercomputers cannot cope with.

Protecting virtual worlds

The proliferation of toxic content and behavior has spread from social media to a new area: the online worlds. Companies are using AI to detect malicious behavior in games and other virtual spaces.

About 3 billion people around the world play video games. By 2025, the online gaming audience will exceed 1.3 billion people. Between 1% and 10% of the North American and European population suffer from psychological gaming disorder. All this raises concerns that children and adolescents are being exposed to negative influences, including inappropriate or aggressive content.

Identifying hate speech online is not a new problem. Facebook says it spent $13 billion on security between 2016 and 2021. The problem of toxic content is quickly spreading to virtual worlds and online games. Games can quickly turn into toxic social experiences in the form of name-calling, cyberbullying, griefing (deliberately helping the opposing team to anger your teammates), and teammates leaving the game early due to dissatisfaction with the way it was going, putting your team at a disadvantage. Anti-Defamation League (ADL) Study revealed that 74% gamers suffer from various forms of toxic behavior. Startups developing technologies to combat toxic behavior are using AI. Spectrum Labs assertsthat its natural language processing (NLP) platform facilitates the process of moderating audio and text content on the 50%, while improving detection of toxic behavior by 10 times. Similar  GGWP entered the market with a service for detecting and monitoring toxic content.

Leading technology companies are buying up startups developing AI-based content moderation systems. For example, in October 2021, Microsoft acquired Two Hat, whose clients include Roblox, Epic Games and Minecraft.

Of course, it is impossible to achieve perfect content moderation. In an effort to avoid platform censorship, the online community is constantly adapting. However, breakthroughs in key areas such as NLP and image-based classification indicate that AI will be at the forefront of the war against toxic content.

Fighting deepfakes

The spread of deepfake technology is becoming ubiquitous. Deepfake candidates run election campaigns, and AI-generated videos spread disinformation about military operations. The scope of deepfakes has expanded from the generation of hyperrealistic images to the creation of voice and video deceptions, including facial reconstruction, when one face in a video is replaced by another. These fakes, based on self-learning algorithms, became more and more realistic over time. The sheer volume of publicly available videos and voice recordings makes it easy to train AI algorithms and create deepfakes, especially for celebrities. Experts say it will become increasingly difficult to distinguish faces, objects and videos created by artificial intelligence from real ones.

In 2022, deepfakes have become widespread in the media, especially in the political sphere. In March 2022, The Wall Street Journal reported that Yoon Seok-yeol, a candidate in the South Korean presidential election, used deepfakes to improve his image among young voters. "Ai Yun" as it was called deepfake version, looked funnier than its real-life counterpart.

In addition to fake news and political disinformation, deepfakes also target corporate clients and can become a major tool for phishing and extortion among consumers.

“What we're seeing is that fake media is getting better and better and better, and the ability of computers to tell what's real and what's fake, or the ability of users to distinguish between what's real and what's fake, is quickly going to zero.”. says Paul England, an engineer at Microsoft Research Labs.

Для того чтобы устранить риск, который представляют для журналистики и демократии синтетические СМИ, вызванные сочетанием новых форм дезинформации с вирусным распространением Microsoft и BBC объединились с Adobe, Arm, Intel и Truepic для создания Коалиции за происхождение и подлинность контента (Coalition for Content Provenance and Authenticity  — C2PA). C2PA — это organ Standards Setter, which will develop end-to-end open standard and technical specifications on content provenance and authentication.

Microsoft is also separately collaborating with the startup AI Foundation. In 2020, the AI Foundation received $17 million in investment to create deepfake avatars, and also announced the launch of a deepfake detection platform, Reality Defender, together with Microsoft. Reality Defender works with, among others, the US Department of Homeland Security and the US Department of Defense.

Truepic has taken a different route and uses cryptography and blockchain-based technology to identify photos and videos. Truepic's integrated Vision Controlled Capture technology detects photo manipulation using a comprehensive set of tests. All images captured with the Vision app have verified metadata, are considered unedited and verified as originals. Truepic is also working with Qualcomm to add encrypted tags to images captured by smartphones based on the Qualcomm chipset. In 2021, the company has attracted $27 million in funding from Adobe, Microsoft M12, Sony Innovation Fund and others.

Last year, researchers at Meta announced that they could now “decompile” a deepfake image—that is, not only determine whether the image is fake, but also analyze the attributes of the AI model used to create the deepfake.

While technology companies are actively developing solutions to address the growing cybersecurity threat from deepfakes, deepfakes will become more advanced and ubiquitous, making it necessary to find new ways to detect and destroy them.

AI in programming

There are many areas where machine learning is gradually impacting the software production process. One of the most important areas is program synthesis, in which a program can be generated directly from natural language, explanations, or examples. Here, one of the most interesting ways is to automatically create graphical interfaces from sketches.

There are several areas where machine learning techniques can significantly improve the software creation process. One concerns the compilation process. In modern compilers, humans define the output, but the compiler defines the order of instructions and sometimes extensive rework. This in turn can have a significant impact on the performance characteristics of the code. Typical optimizations that compilers use to find equivalent and more efficient programs are developed by hand. But researchers have identified areas where AI significantly outperforms what traditional compilers generate.

On the other hand, decompilation, or reverse engineering, is an important step in many security and malware detection processes, where low-level assembly code is translated into a high-level programming language. The key step is to imbue the resulting high-level language with the semantics of the low-level program.

Еще одна область применения ИИ — это тестирование. Хотя оно всегда является скучной и рутинной работой, выявление ошибок и уязвимостей в программных приложениях является необходимым и критическим этапом создания программ. Одним из подходов к их тестированию является нечеткое тестирование, при котором в программу отправляется широкий спектр входных данных с надеждой выявить поведение, приводящее к сбою или другим аномалиям.

AI makes writing code easier and more efficient, not only helping you write programs, but also automating software testing. AI algorithms are capable of translating natural language commands into computer code.

In June 2021, GitHub (acquired by Microsoft in 2018) and OpenAI (in which Microsoft has a $1 billion minority stake) teamed up to launch GitHub Copilot. Copilot, trained on publicly available GitHub data, turns comments into code and can work with multiple natural languages.

Microsoft isn't the only major tech company working in this direction. In February 2022, Google DeepMind released AlphaCode - programs generated by artificial intelligence that have been tested on more complex programming tasks. After evaluating its AI in recent competitions held at Codeforces, DeepMind said its AI performs "around the level of the average competitor."

As for startups, they are more focused on creating software testing systems: automating quality checks and unit tests of code. Major companies in this area include Mabl (supported by CRV and Google Ventures), Authify (backed by Salesforce Ventures) and a spin-out from the University of Oxford Diffblue.

Computer-aided programming is still in its infancy. But the field's rapid development and advances are fueling the increasing use of natural language command-based programming, which enables non-technical users to contribute to scientific projects, close skills gaps, and speed up production cycles.

Multimodal AI

Multimodal AI is a new paradigm in which different types of data (image, text, speech, numerical data) are combined with multiple AI processing algorithms to achieve more efficient results. It is an integrated AI model capable of understanding concepts from multiple modalities such as video, text and 2D images. Its use allows for improved content creation and search. Multimodal AI often outperforms unimodal AI in many tasks.

Today, an AI model trained on video data can be used to predict video content, a model trained on text can be used to predict text, and so on. To move beyond specific media types, multimodal AI research aims to be more holistic, using a single predictive AI model to conceptualize information from multiple data types such as text, 2D images, and video.

For example, in early 2021, OpenAI trained an AI model called DALL-E to generate images based on a text phrase. In January 2022, OpenAI released DALLE-2, which improves the output image resolution of the original model by 4 times.

In May 2022, Google launched Imagen, a text-to-image project that is reported to outperform the OpenAI model in terms of the quality of generated images, as well as the correspondence between input (text) and output (AI-generated image).

Earlier this year, Meta published a paper entitled “Omnivore: A Single Model for Many Visual Modalities.” The paper describes an AI model that, after being trained to recognize 2D images of pumpkins, can also recognize pumpkins in videos or 3D images without requiring additional training for the last two media types.

Multimodal AI is already moving beyond academic labs. Google, for example, is using multimodal AI to improve search. So in the future, a user will be able to take a photo of their mountain boots and search for “Can I use these to hike Mount Fuji?” The search engine will recognize the image, extract textual information about Mount Fuji, image and video data from the Internet, process the information received and provide an appropriate response.

New startups that do not have the powerful resources of information giants are also entering this area. So, in March 2022 the company Twelve Labs raised $5 million in seed funding. The company develops AI to understand the context of both visual and audio data to index videos for efficient search.

Multimodal AI research is poised to move beyond research labs and promises to usher in new content creation and search.

End-to-end machine learning

As commercial AI applications scale rapidly, enterprises are looking to rethink existing data management practices to reduce processing time and increase efficiency. This is quite a difficult task. Moving a project from raw data to a finished AI-generated output is a multi-step process, from data retrieval and data quality testing to model development and post-process performance monitoring.

Hundreds of vendors have emerged in the MLOP (Machine Learning Operations) market to handle different parts of the process. End-to-end learning in the context of AI and machine learning is a technique in which the model learns all the steps between the initial data input phase and the final output result. This is a deep learning process in which all parts of the model are trained simultaneously rather than sequentially. A good example of such a comprehensive solution is the creation of a written transcript (output) from a recorded audio clip (input). Here the model goes through all the intermediate steps of data processing, that is, it can process the full sequence of necessary steps and tasks.

End-to-end machine learning providers combine multiple stages of the AI data cycle into a single product in the form of SaaS platforms. The services of such platforms meet the needs of enterprises that need to quickly and efficiently create their own AI-based systems.

In May 2021, Google opened a platform for developing artificial intelligence applications called Vertex AI. The company promotes the platform as a universal tool for data scientists who do not have programming or machine learning experience.

Another company, DataRobot, has been actively expanding its platform capabilities through mergers and acquisitions for several years. During the period 2019-2021, DataRobot made 5 acquisitions, which allows it to claim a dominant market share of enterprise AI solutions.

An important market trend is the development of solutions with functions that do not require deep applied knowledge. In particular, plug-and-play features are widely used to overcome the lack of knowledge and skills in the field of AI.

Another rapidly growing area is “AI for AI”—the use of AI to automate various aspects of the AI development process itself. In particular, we are talking about such functions as checking data quality or developing components of information models.

There is no doubt that the AI field and large technology companies and startups will continue to offer an ever-wider range of AI services, aggressively competing for share of this rapidly growing market.