Scientist of the Year Igor Farkaš: A neural network recognizes a face or an object, but it may not understand them
Professor Igor Farkaš from the Department of Applied Informatics of the FMPI (Faculty of Mathematics, Physics and Informatics) of Comenius University Bratislava received the 2023 Scientist of the Year award in the category Personality of International Cooperation for significant scientific results in the field of artificial intelligence obtained in collaboration with foreign partners, and for the development of an international interdisciplinary master’s programme in cognitive science. In his research, he works with artificial neural networks, which can be used to effectively solve various problems. He also coordinates the activities of the Centre for Cognitive Science at the Faculty of Mathematics, Physics and Informatics of Comenius University Bratislava and is the main sponsor of the international interdisciplinary master's study programme in cognitive science..
How do you respond to becoming the Personality of International Cooperation?
Even though the prize was awarded to me as an individual, it was made possible by the contribution of several people who have been working with me. I received the award for developing international cooperation with partners in the field of artificial neural networks. I am very grateful to my colleagues from the U.S., the UK, Germany, the Czech Republic, Slovenia and Austria. I would also like to point attention to the international master's study programme in cognitive science, of which I am the main sponsor.
Could you describe the programme?
In 2005, I joined the European Socrates project, which was intended to establish a joint programme of cognitive science, through a collaboration with partners from Vienna, Ljubljana and Budapest. The programme has been available since 2007. Among my closest collaborators are Martin Takáč and Kristína Malinovská from the Department of Applied Informatics of the FMPI. The programme is unique in two respects: in the third semester, student mobility to a chosen university of the consortium is mandatory, and in both years students must hold research presentations at the MEi:CogSci consortium conference. Cognitive science studies the human mind and strives to understand how the human mind works, how humans perceive, act, and make decisions. It also raises difficult questions, like what is consciousness. I attempt to study cognitive processes from a computational point of view.
How do you study these factors?
Simply put, on two levels - one is the perspective of psychology - psychologists use research methods using subjects to develop theories about how our minds work. I study them from the point of view of computer science - the mind is inside the brain, it is a material entity that performs calculations based on the inputs it receives. A person can recognize a specific object because of a calculation: our eyes feed information to the brain, which processes it. We used this method, for example, to study early language development in young children by modeling how words are learned. This was a project on which I worked as a postdoctoral researcher in the U.S.
Is there a formula for calculating all this?
No, it is more about the relationship of perception to its outward manifestations which take the form of action. In other words, it is a mathematical representation that the neural network must learn, just as a child learns early on to grasp a toy or to name it when it is recognized.
How can artificial intelligence be used in cognitive sciences?
Let us imagine a computer program that scans a face with a camera, recognizes its emotional expression and places it in the appropriate category; for example happiness or sadness. The approach so far has been to program rules describing how emotions are expressed on a face - raised corners of the mouth mean that the person is laughing, wrinkled forehead means a frown. The program uses these attributes to evaluate the emotion. However, there is a more modern and effective way of recognizing using neural networks. Instead of giving the system rules which determine the answers, we give it examples of people displaying a particular emotion. The neural network takes those examples from different categories and learns to recognize the relationship between the input and the output— i.e. the correct answer.
How do we obtain these examples of emotions?
Annotated examples of faces that are categorized by emotion can be obtained from publicly available databases. The neural network is trained on these examples until it gives the correct answers. An untrained network just makes random decisions. The training itself means that the network adjusts its parameters to get closer to the truth whenever it makes a mistake. We let it know by telling it the correct answer. This is the task of the training algorithm. All neural networks need to learn and there has been much talk of deep learning in the past decade.
| The TERAIS Project (Towards excellent robotics and artificial intelligence at a Slovak University) |
|---|
The joint research project focuses on cognitive robotics using artificial neural networks and deep learning, which are currently cutting-edge technologies in the field of artificial intelligence. The research will contribute to the development of robots that will be trusted by humans and will allow interaction on a daily basis. International cooperation will contribute to the creation of networks and long-term, sustainable international research partnerships. The project is also intended to promote connections with other key actors of the local innovation ecosystem, such as information technology companies and technology start-ups. It supports joint events, webinars or workshops organized by the students and employees. The project has 4 key pillars:
|
What is deep learning?
It is a method of training very similar to the one used on small networks, except with deep learning the networks are substantially larger and consist of several layers of neurons. Working with neural networks has become much easier than in the past because we now have advanced tools for their implementation. The software is much more sophisticated, programming has become very efficient. Sometimes, just a few lines of code are enough to get the network ready for training. Today, neural networks have several layers, they are deep, and enable more complex transformations, as well as solve more complicated problems. Each layer has its role and is part of a complex process. There is some similarity with the human brain, which also consists of many layers that are interconnected.
Do we know how the human mind works?
We don't know exactly and I don’t think we ever will, in detail, because new knowledge leads to new questions. Researching the mind is difficult because it is an intangible entity which is also hidden from sight. It can be studied by psychology, using experiments. But the mind certainly depends on the brain, and the brain is made of matter. We don’t even have a complete knowledge of the brain. Then there is the question at which level the brain is studied - do we look at the micro-processes going on in neurons, or do we study the systemic level, the interactions between the various parts. In this effort, the neuroscientific view or computational modeling of the processes taking place in the brain can be useful.
You are the coordinator of the TERAIS international twinning project developed as part of the EU Horizon Europe grant programme. What are the achievements over the years that the project has existed?
This is the third application, the first was in 2018. Our current project manager, Daniela Olejárová, is largely responsible for the success of the project, because she helped us to define the project and write the application. With her help, the project got the maximum number of points. Its aim is to develop the human potential at the Department of Applied Informatics of the FMPI. We want to support research and motivate our colleagues, which is why we created a research support office at the department. We have two strong partners – the University of Hamburg and the Italian Institute of Technology, who help us write joint professional publications in the field of artificial intelligence and support out research stays in Hamburg and Genoa.
More than half of the department's employees, as well as PhD researchers and students are involved in the TERAIS project. We want to do better research than before, and the research part of the TERAIS project focuses on the development of artificial intelligence and robotics. I personally am really interested in humanoid robotics, i.e. robots that look like humans and interact with humans. There are already various prototypes that will see further development. The difference between the bodyless, computer-based artificial intelligence and robots is that a robot can act like a human and make a difference, it can interact with people. However, it is crucial to ensure that robots are safe and reliable, and pose no danger to humans.
Can artificial intelligence be harmful?
That should not happen unless something goes wrong with the hardware or software. New forms of artificial intelligence (AI) are able to learn. The question is whether an AI can learn something bad, something that we didn't want it to learn. Let us imagine a system for which it is critical to give correct answers. If we train it well, it works well - for example, it recognizes emotions with a high success rate. The problem is that current artificial intelligence based on neural networks can work very well, but is not entirely reliable because it is vulnerable and not resistant to malicious attacks. If I purposefully modify the image of a human face which is captured by the camera by intentionally changing a few points (pixels), I could cause the recognizing neural network to suddenly give entirely wrong answers with a high degree of confidence. Small errors in the image will make the network think that it represents something completely different. Humans are resistant to this. Our visual system apparently uses different principles and is not so easily fooled.
Do you have an idea of what measures we should take to prevent this from happening?
One way is adversarial training, which means adding malicious data to normal training data, and correcting the network’s parameters until it provides the correct outputs. However, this just solves the symptoms, not the root cause. It might prevent the network from being susceptible to one type of attack, but there are many of these attacks, and new ones are emerging all the time. We can train artificial intelligence so that it would be immune to some specific attacks, but it will fail to recognize others. Nevertheless, I hope that one day we will find a general solution to this problem. Right now, a neural network can recognize faces or brands, but that doesn't mean it understands them. To the network those are still just images which it recognizes on the level of pixels. A human, on the other hand, can imagine what a horse looks like, i.e. humans can work with mental conceptual representations. Current neural networks find this hard to do.
Do you know how we should change the network so it doesn't just recognize images on the level of pixels?
We can take inspiration from the human visual system. Contrary to artificial intelligence, human vision is not a one-way flow of information. Information also travels in the opposite direction, humans form expectations after perception. For example, if I see a dog, I have some expectations about what it should look - it should have four legs and an expected body shape. However, I know it could still be a dog even if I don't see those four legs. Neural networks also have an attention mechanism - if the network recognizes an object, it might be because it pays attention to those parts of the image that are important for correct classification. For us, the background of an object is usually not important, but current neural network models classify the object in the image partly because of the background, which they consider equally important.
Radka Rosenbergová