Once, graphics cards "speeded up" the game. Now the industry will move into fifth gear thanks to AI machines

We talk with the AI ​​technology specialist about the fourth industrial revolution that is taking place before our eyes. Artificial intelligence will play a key role in it.

Dariusz Dobkowski worked for years in American corporations, dealing with the development of business in the segments of storage and data centers. For two years, he has been leading the sale of advanced solutions from the Huawei offer, where he develops, among others AI department. Builds an ecosystem of partners, developers and partner universities. That is why he is the perfect person to talk to about the state of artificial intelligence in Poland and in the world.

Karol Kopańko, Spider's Web: Where is AI now?

Dariusz Dobkowski, Sales Director at Huawei: This question should be answered by adding some context.

Moore's law is no longer valid. The computing power of general purpose processors cannot be doubled in a short unit of time. We are limited by physics. In the past, when manufacturers released a new generation of processors, we made a technological leap. This is not the case now. So there is a switch to narrowly specialized systems.

And AI needs huge computing resources.

Exactly. Historically, such power was available only in data centers, but it was associated with very high costs. Later it turned out that graphics cards also do great in these applications, but the demand on the gaming and cryptocurrency market has pushed GPU prices. That is why specialized systems are currently appearing, whose task is to accelerate calculations on neural networks.

So we are dealing with a similar evolution that the computer market went through nearly a quarter of a century ago, when 3D accelerators appeared, accelerating the generation of 3D graphics.

Yes, you can use this comparison. And just like in the case of games - here you have to write the appropriate software to be able to use the equipment.

So Huawei creates chips that specialize in AI so that developers can later train their neural networks on them.

Exactly, we currently want to create an ecosystem that will develop AI. We will invest EUR 100 million in Europe over the next five years to achieve this. Goals are clearly defined. We want to attract 500 new partners and 100,000 developers who will create AI solutions on our platform. For this we want to work with hundreds of universities throughout Europe. We already have the first partnerships in this field in Poland - including with Warsaw University of Technology.

It's an ambitious plan. Why should these three elements of the environment - universities, partners and developers - look exactly at Huawei?

Huawei does not offer end-to-end solutions, so building an ecosystem is part of the strategy for us. We bring to it not only our hardware, but also funds for the development of technological thought or marketing activities. We want to create joint solutions and laboratories, give access to our competence centers and listen to the needs of the environment.

Suppose I represent a small company that produces cameras - because AI is great at recognizing objects. How can I benefit from cooperation with Huawei?

A great example. A standard camera partly supported by AI copes with the recognition of about 30 objects in the frame. Let's imagine that we want to check if all customers visiting a given store wear recommended face masks. Our chips can increase performance by up to seven times, so we will be able to check up to 200 people in this respect. It is thanks to our tiny solution - Atlas 200. It is the size of a credit card, so we dedicate them to small devices that have to work on batteries for a long time - cameras, drones or robots.

Why mount the chip directly in the drone when you can send the signal from its camera to the data center?

A progressive trend in AI is edge processing, as opposed to processing on a central server. Instead, distributed entities process information themselves and can make decisions.

Let's take the example of a drone monitoring forests in remote areas. Transferring large amounts of video is inefficient and expensive. Instead, the drone can independently analyze the material and send only the most difficult, ambiguous freeze frames. Such a drone could detect fires and only confirm requests from a central server. We have just implemented such a solution in very large forest areas.

I understand that this will help generate savings in the field of wireless communication and will significantly speed up the entire process, because the calculations are generated directly in the device?

Exactly so - in this type of applications the ratio of performance to power consumption is crucial, so that the given device can perform the tasks set before it and at the same time work for a sufficient time.

What about training the model? My guess is that it happens in data centers.

Definitely. Here we have the Atlas 800, which in turn counts for pure performance. It is on them that the deep learning process takes place, i.e. the most complicated stage of training, when the model learns to draw conclusions similar to the human mind.

For comparison, Atlas 200 can process 16 high definition video streams simultaneously, while Atlas 800 has 512 channels. Its computing power can also be increased by using the Atlas 300 cards installed in it. And - to mention all the elements of the Atlas family, there is also the 500 model, it's an AI-in-the-box, and therefore an independent unit that can be ideally suited to companies or stores, without the need to install a dedicated server.

In turn, developers starting cooperation with us can use the Atlas 200 Developer Kit. This is a starter kit that allows you to familiarize yourself with the hardware and libraries.

So what does the adoption of this technology look like in Poland?

In Poland, we currently have about 300 companies that prepare solutions based on AI, but interestingly, 75 percent. their revenues come from abroad. Polish customers are still sniffing at AI and checking what benefits it can bring them. Unfortunately, economic issues limit us, so education is needed all the more because investing in AI brings savings in the long run.

What investments can it be?

One should not ask what AI can do, but rather what we want it to do. We are slowly approaching the moment when we can already say that it is enough to point out the problem and it will be possible to program it accordingly.

One of the most unusual applications of artificial intelligence from Huawei, which I encountered, was the analysis of the spectrum of the northern lights, on the basis of which the algorithm composed a musical piece. Later it was even performed by the philharmonic orchestra.

And when it comes to more practical issues.

Of course, it starts with repetitive activities. I myself saw a production line of smartphones with a performance of tens of thousands of copies per day, where only a few people worked. Adapted machines adapted to the rest of the process.

AI already uses logistics (package monitoring) or financial (risk analysis or KYC) business. We also see many applications in medicine, where we can help doctors detect or analyze the progress of the disease.

So we are talking about helping doctors, not replacing them. At least for now.

I recall the results of comparative studies, which showed that specialized algorithms are already effective in diagnosing bone diseases in children at a level similar to recognized specialists in this field. Here, AI will make a big career. I think that everyone can remember the situation in which they had to wait a long time for analyzing the X-ray, because the specialist was not available at the time. Now the photo can examine the algorithm at first, and the man can confirm later.

And if we are already in the medical field, then in China AI is already controlling vehicles that during the COVID epidemic distributed hand disinfectant liquid to the estates and allowed it to be purchased in a non-contact way.

One of the examples of our projects that touched me a lot is the StorySign project. It is an application that helps deaf children from birth to learn how to read. The child can read the texts through the screen of the phone, which in addition to the text being read appears with a nice avatar showing the word in sign language. This is one of the projects that changes the lives of others for better and easier.

We can also mention the implementation using drone image analysis to examine the condition of high voltage lines. Earlier, it was necessary to send a man to risk his life on poles, high above the ground, and now it can be done much faster and more accurately using a small device supported by AI technology.

However, we are constantly talking about the applications of AI as a curiosity. What is the biggest challenge in this context that stands in the way of wider adoption?

It starts with money, because you have to invest in training the model, and here the cost can go into thousands of dollars. However, it seems that companies need more business imagination to be able to see inefficiencies in which AI can perfectly fit.

It seems that AI is the future, and whoever gets on the train with the inscription revolution will benefit the most.

Definitely. It is indeed the fourth industrial revolution that is happening right before our eyes. Only two years ago, only 4 percent. companies have admitted that they use AI commercially. It's not much. It is different if we look to the future. By 2025, as much as 80 percent I want to have AI at home. If this really happens, then we will have a real boom for AI services.

Huawei is ready to support all actors of the AI ​​ecosystem. We have global reach, so we can promote an interesting solution from Poland around the world.

Once, graphics cards "speeded up" the game. Now the industry will move into fifth gear thanks to AI machines