The year 2002 was - seen through today's eyes - basically digital prehistory.
Yes, but the constant change in our industry is probably the main reason I’ve stayed with one company for over twenty years. I've been through so many different positions and worked on so many different projects that it never became mundane or routine. I was always getting into things that were new. I’m constantly surrounded by young people and innovative technologies.
In the past, you worked a lot on Industry 4.0, for example, and today you’re working on projects for the VPT4 team, which is focusing heavily on artificial intelligence and machine learning. What are you working on at the moment?
We’re in the proof of concept phase of a large-scale project that will help make motorway repairs faster and cheaper in the future.
What’s does that kind of project involve?
For more than five years, VARS has had a vehicle that it drives along individual sections of motorways. And it’s equipped with cameras as well as laser sensors. They not only enable it to detect major road damage, but also find micro-cracks in which some damage is spreading. Those images were then handed over to a team of people who manually inspected and graded them on a damage scale of 0 to 3. They then prepared an analysis for the Road and Motorway Directorate on how and where to repair the carriageway to ensure the best price/performance ratio. Our job was to use artificial intelligence so that a human wouldn't have to do it. And, at the same time, ensure even greater precision. We're now at the stage where we've gone through the images with the customer, evaluated them, and confirmed that we're able to evaluate it better and more accurately than when people were doing it.
How does artificial intelligence "learn" in the first place?
It is not entirely unlike how people learn. These days, artificial intelligence is more or less just a lot of advanced statistics combined with algorithms that try to approximate how our brains work. In practice, with this particular project it looked like we first took the photos and marked the ones containing errors. We call this principle "learning with the teacher" - you need to tell the system that makes up the neural network and does the computing what the result actually looks like, what the input is and what the output might be. That's why we cut everything up into tiny tiles. This is preparation of the data, without which you can’t train the neural network. The second step is that you divide the data into training, validation, and testing. It works similarly to how we learn a new language or some other discipline. By using the principles of neurons and statistics, the algorithm derives results from a large amount of data. Once you've trained it, you pass it on to the next part, validation. There, you compare the actual result with what the neural network can do. And if it intersects well enough and you get an interesting level of probability, you report that it has the requisite accuracy and start applying the result in production.
How many of these trials are needed before the result is satisfactory enough to pass on to the customer?
We were relatively quick and efficient in this project. In similar types there are usually dozens of iterations. We were able to input the right prompts and the system learned quickly, which is why we got below twenty or so before we trained the neural network.
Forgive me for being a bit of a layman - but generally one gets the feeling that AI is mainly the domain of the biggest tech players, who invest huge millions of dollars in its development. How can a Czech company be competitive?
We don't have this model based on having a system or software that we grab and start applying. Our competence is built on the team's ability to apply the capabilities of the various platforms in use today. There are lots of people in our VPT4 team with a mathematical rather than an IT background. We’re the ones who create specific solutions for individual customers, but we don’t have a universal and prefabricated product. The strength of the team is not that we have a product and a single path that we repeat over and over again. However, we do have a methodology that allows us to grasp different topics related to the processing of unstructured data, such as image, audio, and evaluate it on the basis of that particular case. We don't completely believe that someone has a trained neural network that can do everything, and so things will be all rosy and great. Because often the basis is that the data is specific, and training needs to be tailored to that. What makes us unique is our ability to apply artificial intelligence and machine learning capabilities for the benefit of specific customers.
Does it help in any way that you’ve been in the IT infrastructure business for a long time?
Definitely. We have a huge computing cluster for this kind of project, which is a huge advantage, as things like this would be extremely expensive in the cloud. We have a few dozen graphics cards that allow us to quickly describe the task, create the output, and get the neural network trained fairly quickly. We have good infrastructure that is capable of computing on neural network graphics cards. Everything we need is available through our own secure data centre in Lužice near Hodonín.
What's next for the project to check for and quickly fix potholes on roads?
Once the concept is completed, it will become a classic project. We’ll deploy it in a subsequent project as a production issue that will be commonly used for image evaluation.
What does the VPT4 team do besides image recognition?
We do a lot of NLP (an area most of us who have tried GPT chat are familiar with), which is text recognition and natural language processing and text generation. Equally attractive for us is DataMesh, which is the principle on which today's modern data warehouses capable of quickly processing big data are built.
And what's next for AI?
It makes sense to me not to use humans for routine activities, and that's where I see AI as an extraordinary asset, since that's an area for which it’s well suited. But otherwise, it's a huge debate. So far, we are the ones who decide what the machines should do. The problem arises the moment it starts to be the other way around. We can already see around us how effectively AI can be abused, such as for disinformation and the manipulation of public opinion. But I don't want to be pessimistic - I'm confident that the positives will far outweigh the negatives and that we’ll oversee the whole process of defining and applying the rules of how AI is used.