Today in the world of the development of science and high technology there has been a catastrophic substitution of concepts: for science it is issued that science is not in any way approximated, scientists call programmers and engineers, science calls the solution to the simplest engineering problems. In the information space, the role of fundamental science in the development of technologies is clearly underestimated. Many people forget that the iPhone’s touch screen is not the iPhone’s touch screen itself, but the implementation of the ideas of fundamental research on the semiconductor heterostructures of our compatriot Nobel laureate J.I. Alferova. Google (or Yandex) maps are not just maps in a mobile phone, but the embodiment of fundamental research in the field of computational geometry. And by the way, the Avatar movie is also 99% computational geometry. Machine vision, neural networks and artificial intelligence are no exception: this whole complex works well and correctly only because fundamental research is the basis. A fundamental approach to development, on the one hand, is the key to the right vector for the development of the industry, and on the other, it demonstrates various garage startups that flooded the market, which fundamental science always gives a clear advantage.
We are scientists. We are scientists and do not hesitate at all. Our startup (which in fact has not been a startup for a long time, but a rather well-functioning business, known both in Russia and abroad, developed exclusively due to the developed technologies without any investments from the state or funds) is about the very hype , which everyone is hearing today: computer vision, machine learning, artificial intelligence, neural networks - in general, the whole set of words bullshit bingo, the use of which, according to many startups, makes the project obviously successful. This, of course, is not so. It is important whether you understand the essence of the issue. That is why many starts from the up position very quickly find themselves in the down position, and then in the out position. Because not scientists.
We rarely go to meetings of startups, visionaries, and evangelists in the field of artificial intelligence. The fact is that now everyone has learned to make beautiful presentations. Who did not learn - found a contractor. Huge resources are invested in the promotion and promotion of technological solutions (first to attract investment, and then to justify the activities of a weak team), for which there is neither novelty nor efficiency. In a beautiful wrapper for presentations, infographics and animations, especially when fashionable and obscure scientific and technological terms are woven into them, you can always hide the lack of a real sense of activity. For many presentations there is no science. It is a void wrapped in a beautiful shell. This is a bait for investors who peck (or pretend to peck) on shiny glass and give money it is not clear for what tasks. In reality, few want to understand the scientific component, most prefer to pay attention to the outside of the “pitching”. What to do - presentation economy in action. And we are not talking about that. We are about science.
We go to scientific conferences. For example, we were recently in Australia at a document recognition conference. Russia - and not only it - is a country with an increased level of development of the bureaucracy. Some, suffering, collect pieces of paper, others with no less suffering work with them. Ministries announce tenders for digitalization and automation of routine processes, quite serious companies-executives appear, offering their approaches in this area. From the high stands it is explained that this time it is time to replace not only the eye and hand, that we are talking about artificial intelligence that can understand documents in a meaningful way. And in this context it is very strange that only we and Abbyy were at the leading profile conference from Russia. We did not see representatives of the flagships of digital transformation and participants in the state program for the development of artificial intelligence. It turns out that in the field of understanding documents in Russia there are no more scientists?
We have just returned from Amsterdam, where the ICMV , a scientific conference on computer vision, was held. There, we did not at all set the goal of promoting the success of our company as a business structure. We told the community of professionals what fundamental problems our research team is working on. For our young scientists (it is precisely the young employees who offer and develop bold ideas in our field who are the keynote speakers at the conference) this is an opportunity to immerse ourselves in scientific activity, to develop ourselves scientifically and professionally.
We are sure that for a startup to “take off” and become a developing business, it must be based on science, from which a working technology will grow, which is the “engine” of a service or product that is in demand by the market and meets its requirements.
In our area - areas of recognition (these are identification documents, various profiles, bank cards, tables, barcodes, images) - the technology and science underlying it is the foundation. The price of the error is very high, and that is why we have very high requirements for the algorithms. It's like a plane - it either flies or does not fly, there is no third. If it flies badly, it means it does not fly at all. And just like in aircraft manufacturing, the algorithm is based on science, a serious fundamental science, which grew out of the backlog of our compatriots who were at the origins of our, home, artificial intelligence. Many are now engaged in recognition, creating a market for these services, which is still in its infancy. It is not fully formed, despite the fact that analysts predict the global volume of this market to 16-17 billion dollars by 2024. But we see very few colleagues at scientific conferences. Even fewer are those published in scientific journals. Now we can hear how the slippers of indignant startupers flew into us, who are surely convinced of the futility of conducting research and writing articles. “We are doing business, not some kind of theoretical science!” This is precisely why many recognition systems on the market work frankly poorly (despite beautiful presentation videos and advertising slogans), they are mistaken, they require the participation of a human operator (or an entire factory of these operators) , which will examine the fuzzy image and adjust the decision of the machine with its own tired handles. But the task is to make it so that the recognition process is completely automated. In our case, this does not mean to eliminate the error completely, but to make the machine make mistakes many times, tens and hundreds of times less often than humans.
We are constantly working to improve our own technology. This is a process associated with a lot of theoretical work and serious research. In technology, we improve accuracy, speed, flexibility, make the technology more “light”, less demanding on hardware, less energy-intensive. This is our “green AI” (green is not in the sense of “immature”, but in the sense of “green”). We understand that environmental technology is a trend. And potential customers in the world will cling to this trend. And we got an understanding of this trend precisely from participation in scientific conferences. As the Scandinavian proverb says, “When you cut the forest, do not forget to sharpen the ax in time.” Participation in scientific conferences for us is just the process of sharpening an ax. Technology cannot grow from scratch, understanding how it can be improved does not appear immediately. The scientific community is fundamentally different from the community of startups, investors, blockchain analysts and technology visionaries. It is not enough to show a beautiful presentation here. If there is no thought and novelty in it, they will peck. They will eat and bury. Yes, and arranged at scientific conferences, everything is a little different. We show there not the final result, but what leads to it, describe the methods and approaches, trample a glade for ourselves, presenting the world the results of our own research. It’s not enough to show what you have done, you need to explain how it works.
Now, many startups in the field of AI are such peculiar auto repair shops where they do not repair the car, but replace the faulty nodes. Something does not work - we don’t fix it, we immediately change the module. This is easier, especially when you don’t know how it all works, and what you need to do to fix it.
We are scientists, and quite confidently show: the technology based on deep fundamental science works better, faster, more confidently, more reliably than the one where of all the technologies there is only a beautiful presentation. It breaks less, consumes less energy, works in the snow, in the heat, at night and in the morning, since we know what kind of neon is in its “inside” , in which place it is necessary to “ pull the little girl ” and hit the tambourine.
Our company employs more than 50 developers, each of whom destroys the thesis of the lack of demand for scientific work. Our team includes both established scientists and those who are just starting their career in a scientific career. We emphasize that these are not just programmers who deal exclusively with code. Each of them is an independent scientific unit, from which the scientific team is formed, the flexibility and variability of the solutions used is ensured.
And all because we go to scientific conferences.
At the international conference on computer vision (ICMV) in Amsterdam, we proposed our fundamentally new approaches to the development of neural networks and recognition technologies, which are aimed at reducing the carbon footprint and minimizing environmental damage from the use of new technologies. The main emphasis was placed on the optimization of computational algorithms used in hardware and software systems, which should reduce the energy consumption for training and the functioning of neural networks on a global scale.
Today, the sustainable development of our planet is becoming the main agenda in the reports of world organizations concerned about the future of the Earth. And this is not only Greta Tunberg. A person’s place on the planet in many respects depends on how much we can balance the development of technologies, the growing appetites of global corporations associated with them, and responsible attitude to the environment. Possible ways to minimize the harm from rapidly developing technologies of the last decade are largely associated with reducing energy consumption and finding the most optimal and energy-efficient tools for solving modern technological problems.
Our arsenal has neural networks and it is our responsibility, scientists, to make them work quickly, efficiently, and correctly solve the tasks assigned to them. From the point of view of “green artificial intelligence”, the issue of a breakthrough is not to create a large, powerful, huge neural network - a sort of blue whale with its surprisingly large neural network, but a fly, with its very modest “computational” capabilities, but its ability to quickly and energy efficiently solve the necessary tasks. Such a neural network should have a completely sky-high specific productivity. The right choice of algorithms allows you to spend 1000 times less resources on image recognition tasks than some equipment manufacturers claim today.
Today we understand that in our area, increasing the capacity of devices for performing operations is a one-way ticket. And the task of scientists today (with a responsible attitude to development) is to make the technology so that it loads the devices as little as possible, does not force the processor to warm up to the temperature of the boiler, does not eat energy, like a “gastro-unsatisfied cadaver”. And then this technology becomes user friendly and at the same time high-tech.
For this to become a reality, completely unobvious “bundles” of completely different fields of science are required. Is it clear to many that shifting deep learning technologies to tropical algebra can dramatically reduce the number of transistors in neuroprocessors while maintaining expressive power? Or is it that the presence of Radon layers in a neural network allows it to economically describe projective invariants, which, in turn, determine the internal structure of images of the real world?
Participation in scientific conferences is, first of all, an opportunity to compare watches with world leaders, to see their real competitors (yes, real competitors also go to scientific conferences, because they are also scientists), to see where technologies are moving first-hand, and not in retelling of visionaries and evangelists, who for the most part have nothing to do with real science, but only learned to put fashionable words into relatively competent sentences. Without this understanding of trends and trends in fundamental science, no movement forward is possible. Technology without science “will not take off,” or it will already be obsolete in advance.