“The question about the future of artificial intelligence should be looked at in a broader context: Where is this all going? Why do we need it? Do we do that only to make money or to improve the quality of life?” says Jarosław Protasiewicz, PhD, director of the National Information Processing Institute, in conversation with Maciej Chojnowski
Maciej Chojnowski: Much criticism is leveled at artificial intelligence these days. Gary Marcus claims that deep learning, despite numerous achievements, cannot be regarded as the ultimate AI paradigm, as it still unable to provide cause-and-effect reasoning. Jeffrey Funk argues that the hype about AI is fueled mostly by marketers or investors and that we are dealing with a speculative bubble similar to the Dot-com Bubble, which burst at the beginning of the 21st century. You are a long standing AI practitioner. What is your take on those phenomena?
Jarosław Protasiewicz, PhD*: I tend to agree with such opinions. Although I don’t think this bubble will be the same as the one we saw 20 years ago. Everything will depend on whether the money is there or not. I believe there is but it is not as big as some might expect.
Nowadays, many corporations “refresh” their portfolios. What was once called data mining is now referred to as artificial intelligence or machine learning. We are following the latest trends. In a few years a new slogan will be coined and many people will readily use it for their own ends. All stars earning a living by attending conferences will be saying different things then. But, surely, something will remain and scientists will continue doing their research. Although the scope might be different.
But we still don’t know what the world will look like in a couple of years. Maybe, due to the epidemic, the paradigm of continuous growth will be dismissed? The question about the future of artificial intelligence should be looked at in a broader context: Where is this all going? Why do we need it? Do we do that only to make money or to improve the quality of life?
The problem is also to verify to what extent AI is really effective. Funk took a closer look at enthusiastic reports prepared by consulting companies to check the real accomplishments of the American AI start-ups with the biggest subsidies. It has turned out that extremely optimistic forecasts have proved only partially accurate. At least for now. So is it possible today to identify the areas where AI is really effective?
Artificial intelligence, in a broad sense, has been a part of industry for a long time. To give you an example, think about different models pertaining to failure prediction or automatic control of industrial processes. Obviously, a great majority of such solutions in the field of industrial automation rely on proportional-integral-derivative controllers. You also must remember that the AI model training boils down to calculating first derivatives of an objective function. We were taught that in the high school.
There aren’t too many truly effective AI solutions. More often than not, AI is just an add-on which, technically speaking, doesn’t make much difference. But it may be key to sell a product.
But there are other important solutions, e.g. natural language processing or translators. In the past, they were based on statistical models. Today, recurrent neural networks reign supreme, and they work like a charm. Other solutions include speech to text and automatic speaking. And there is of course image processing. It has already been applied in autonomous vehicles. The big question is whether self-driving cars will be able to operate in an open system and not only in an isolated testing environment. In my opinion it is not possible. But new methods are being developed every day.
It looks like there is a lot of examples.
Yes, but there aren’t too many truly effective AI solutions. More often than not, AI is just an add-on which, technically speaking, doesn’t make much difference. But it may be key to sell a product.
We definitely need new solutions. Development was possible not because of artificial neural networks (the approach which has been known for years) but because of Nvidia and its processors and because of increasingly better programming libraries. Today, anyone who understands basic concepts can create an AI model. The fact that most models are designed out of line with professional standards is a horse of another color.
What does this incompatibility consist in?
The results are provided on the basis of training sets. Validation and test issues are neglected entirely. The core principle is not there. In consequence, such results are often incorrect.
To make things worse, various theses and papers surprise us with conclusions like “We have improved the efficiency of our model by 0.5 percentage point”, which actually is a lot because sometimes the increase is of 0.1 percentage point. We should ask ourselves a simple question: Why are we even doing this?
Obviously, research done by a scientist can be art for art’s sake as the sole purpose of the scientist is to get better. But ultimately, if the scientist doesn’t invent a new method, stays focused on optimizing the old one and improves the accuracy by 0.5 percentage point, then what is the point of all that? And what if we simply regressed and accepted to lower the result by one percent? Maybe someone who would use that model in the future wouldn’t care about it at all?
I think we are getting overwhelmed by a techno-optimistic discourse. Everyone keeps telling us that AI is the way to optimize different things but in fact no one knows what the real benefit will be. Are most of today’s AI projects implemented, despite being inefficient, only to serve marketing purposes and to increase the value of tech companies shares?
In most cases AI models and algorithms bring slightly better results. But are they really necessary? They are certainly needed to sell a specific product. The product is advertised as “smart” because people fall for it. Even though it may not contain any algorithm or have a tiny bit of it.
On the other hand, there are many useful products. Just to give you an example: in Thailand I was able to communicate with a Chinese at a bonfire party with my iPhone translator. It was super rough but it was still better than nothing.
At university I specialized in robotics. I remember that already 20 years ago attempts were made to implement algorithms that would allow for better machine control. But the problem was always pure mathematics; for example, you needed to calculate if there was a risk for the robot’s arm to resonate. That was absolutely necessary.
So, today, even though an artificial intelligence algorithm may only be an interesting add-on making a slight difference, the most important thing is to create a useful product.
That is supposed to be the key criterion.
Some time ago I went to the US to attend a conference on brain-machine interfaces. An inventor from Berkeley came up with an idea to build a simple EEG controlled robotic arm that would make possible for a paralyzed person to drink from a cup without being fully dependent on their caregivers. He said that the basic version of his invention had to cost one thousand dollars, a sum affordable to disabled persons.
I don’t know if he has succeeded. Although I heard dozens of other people giving their speeches, his presentation is still engraved in my mind. Utility: irrespective of whether we are talking about AI or not, a solution has to improve the quality of life and be affordable. That is paramount.
Many people criticize research and development in Poland. Some claim it is partially due to the transformation in 1989, when many laboratories closed down. Or maybe Polish entrepreneurs are not interested in innovations?
It is true that after 1989 a lot of R&D labs closed down. In 1998 I graduated from my robotics studies and I couldn’t work in my profession as robotics in Poland had disappeared. It is back now, but then I had to take on computer science.
Universities don’t know how to adjust to the needs of entrepreneurs. Entrepreneurs don’t want to collaborate with them because of their archaic organization. That is the main problem.
I see entrepreneurs in a rather positive way. They look for good investments, only that in Poland we lack big capital and extensive international contacts to create complex supply chains. We are focused on simple products.
On the other hand, simple products benefit the economy the most. The super innovative ones do not translate into a huge national income.
And what is your view on the collaboration between science and business in Poland?
Universities don’t know how to adjust to the needs of entrepreneurs. Entrepreneurs don’t want to collaborate with them because of their archaic organization. That is the main problem.
Another problem is the fear of collaboration; in Poland we have to comply with the public procurement act and people are afraid of being accused of a scam and having their business ruined by someone else who is ready to do everything to earn one zloty more. That is why entrepreneurs are not willing to cooperate with institutions. They prefer to collaborate directly with specific persons.
But I wouldn’t say we are doomed to fail either. An increasing number of young people have been setting up their own companies and trying to do something in collaboration with universities and bigger companies. With that in mind, we have to think about what entrepreneurs really need. And all they want is to do business safely. Of course, there are also aficionados who want to do what they like. Mutual trust and contacting business partners in person also play an important role in the relations between science and business.
I have recently listened to the debate regarding that issue held by the Coalition for Polish Innovation and the Foundation for Polish Science. A representative of the business sector said that entrepreneurs have to learn how to talk with researchers and that researchers have to accept the fact that entrepreneurs need specific solutions. Do you think there is going to be a change, considering the requirements introduced in the Constitution for Science?
I am an optimist. I believe both sides are willing to cooperate. The business sector needs to learn how to collaborate with universities but universities have to be open for discussion.
I remember some people from universities coming here to perform a specific task. Unfortunately, in most cases, instead of a chair we were given a design of a hammer. Just because those people had a “holistic” approach. This example shows how divorced from reality many people from universities are.
When I started my doctoral studies I said that my algorithm had been implemented in a power distribution facility and that it had been performing very well. But the professors were interested in the publication only. The implementation was irrelevant. Fortunately, some academics have changed their approach.
It is said we are witnesses of a new arms race. Only that this time it’s not about missiles but about technologies, although autonomous weapon is an important issue as well. Where do you think Poland is on this map? Do we even have a chance to find our place in that race?
Poland is not the main player. We do what others allow us to do. We do not develop industries that should be developed for obvious reasons because I think we can’t. In fact, many countries can’t do that either.
Do we have any chance? We sure do, especially now when some claim that many European companies may shorten their supply chains and move production to somewhat cheaper Central and Eastern Europe. Maybe we will be able to grab a bigger piece of the pie.
Does it mean that the Polish artificial intelligence development policy should identify specific areas and develop them according to a plan or leave it to spontaneous development of the market?
We have to set strategic goals. Also, the Polish state has to assess strategic directions based on experts’ opinions and not on somebody’s arbitrary decisions.
The invisible hand of the market has been compromised. The state has different tools at its disposal to stimulate growth, but it is also responsible for physical, economic and health safety of its citizens. So when it comes to defining various development strategies (also concerning artificial intelligence) and stimulating that development, the state is obligated to provide its citizens with space for a more comfortable, better and safer life. If that condition is met, there is a chance for all that to work.
This mindset can be clearly seen in the Integrated System of Information on Science and Higher Education 2.0 (POL-on 2.0) and the RAD-on integrated information platform, both being developed by the National Information Processing Institute. RAD-on is a flexible tool that will make gathered data available not only to the institution that requested for them but also to other interested parties. What is the role of artificial intelligence in those systems?
POL-on focuses mainly on the use of data searching tools. It’s mostly about natural language processing technologies. As far as RAD-on is concerned, we are now at the stage of building business intelligence based on traditional tools, but are also planning to use more complex machine learning models.
We have to set strategic goals. Also, the Polish state has to assess strategic directions based on experts’ opinions and not on somebody’s arbitrary decisions
We want those data to be widely available so that entrepreneurs could use them to create new services. We have established cooperation with a company from the Czech Republic. And they really do download the data and use them.
What kind of data are those and why are the Czechs interested in them?
The company provides tools to choose and match experts from various fields and wanted to know what the Polish potential was like. We have signed a cooperation agreement regarding new projects and we will try our best to expand our business.
It is noteworthy that we have drawn the attention of a company from the Czech Republic and not from Poland.
Maybe in Poland we lack concrete examples to illustrate real benefits of new technologies? Germans have decided to adopt that model to educate entrepreneurs; your discussion about artificial intelligence with a farmer will be different from what you have to say to a bus manufacturer.
Yes. Examples are very important. But you also need open-minded decision makers. Our institutions must be internationalized. We must attract professors, doctoral students and postdocs from other countries. We have to cooperate with different companies, not only the Polish ones. An extensive network of collaborating entities is a key to success.
Let’s talk about the concept of data-driven economy. Does it really hold water or is it just marketing?
There is a theory that whoever has data, has power. But I’d rather say that power is held by whoever has access to verified information and not only data.
There are two approaches to how knowledge is created. The first one is very popular these days: we have data we can use to get something relevant. But before that happens, we need to ask ourselves whether such data are complete and whether they present a true image of what is going on. There is a risk of getting incorrect data and using them to create false knowledge.
It is not fully understood how artificial intelligence works. That’s why accuracy and explainability are of paramount importance if we want to use AI in situations influencing human lives.
The starting point of the second, older, approach is expertise. Sometimes you have to use a top-down model and ask a group of wise people how things are. it is not always necessary to click your mouse one million times. In some cases a data-based economy makes sense, provided that data are relevant and reliable. However, in many cases it doesn’t make any sense at all.
In “Rebooting AI” Gary Marcus claims it is necessary to include expert systems in the today’s way of thinking about AI. He provides several examples of innate knowledge: few days after they are born, ibexes are able to climb up and down walls of rock without having a chance of doing that before. So maybe some things should be programmed in a top-down fashion. Does that kind of merger of AI paradigms make sense?
Definitely so. Generally speaking, I would opt for supervised learning: to include expertise in the models or to correct the models based on that expertise. Of course, experts may not notice some things and data may lay the groundwork for new knowledge. Such knowledge should also be verified by experts. Those two approaches must be complementary.
Today Europe calls for ethical artificial intelligence. The problem is that the approach to data on the old continent, in the United States, and, especially, in China is completely different. On the one hand we want to protect the right of the citizens to their privacy and being anonymous, while on the other hand we wish to develop tools based on big data sets. Do you think it is feasible?
Well, technology is supposed to serve us and not to be a goal in itself. If anyone doesn’t agree to have their data processed, they should have the right to say “no”.
But, as you can imagine, there are areas where data are processed irrespective of whether someone has agreed to that or not, because this is how we ensure public safety or track down criminals. As far as GDPR is concerned, I think we were too restrictive. Or maybe it was we who created some of the problems in Poland?
So how to develop new technologies in line with ethical guidelines?
As far as gathering data goes, we need to find a golden mean. The second thing is explainability of artificial intelligence. This is extremely important.
When I dealt with industrial automation and created the control system for a process, there was no room for mistake. If I did something wrong, it would be a disaster. In engineering, everything had to be accurate.
In computer science, software has a lot of errors. It is developed quickly and nobody bothers to be very precise. It is still a workshop. The same goes for artificial intelligence. It’s not fully understood how it works. That’s why accuracy and explainability are of paramount importance if we want to use AI in situations influencing human lives.
What is interesting, originally, my master’s thesis was supposed to be about controlling an infusion pump and dosing medicines to patients based on their temperature and intracranial pressure. I was getting ready, I was scheduled to participate in surgical operations. But then I was asked: “And what if you are successful in developing your solution but the patient dies? Who will be responsible for that?”
That’s a fundamental question…
… which was asked in 1997 before I took on my master’s thesis. As for the answers, little has changed since then.
Jarosław Protasiewicz, PhD – director of the National Information Processing Institute*. He defended his doctoral thesis at the System Research Institute of the Polish Academy of Sciences. A lecturer in the Warsaw School of Information Technology. For many years, he was the manager of the Laboratory of Intelligent Information Systems, which is the biggest laboratory in the National Information Processing Institute. His long-term professional career has been connected with the R&D area. His interests include artificial intelligence, software design, statistics, time series forecasting, as well as text and web mining. Currently, his focus is on managing projects with the use of agile methodologies, creating and developing software, machine learning, bioinformatics, and big data.
**The National Information Processing Institute is the publisher of the portal sztucznainteligencja.org.pl