Scientists working on AI in Tartu

Photo: Aivar Pau

Tartu has become the home of an international high-level working group of scientists that can now, after four years of hard work, create artificial neural networks that function similarly to the human brain, as well as train and make them cooperate.

Syria, Switzerland, Spain, Germany, and several other countries, including tiny Estonia – these are the countries of origin of young computer and neuroscientists the primary interest of whom is to study the human brain with the help of artificial intelligence systems they have themselves created.

Work is being done on two fronts: on the one hand, the scientists are learning about the human nervous system; on the other, researchers are trying to translate it into the language of mathematical algorithms and artificially copy it.

“Interest in the brain and in artificial intelligence is the main thing that has brought us together,” said head of the working group, professor of data science Raul Vicente from Spain. “Brought us here, as the University of Tartu computer science institute has a lot of people researching these things on a high level.”

Cooperation between several European universities happens here – in the so-called Paabeli building accessible through a gateway opposite the main building of the university. Recordings of brain activity are sent in from the University of Lyon in France and studied with the help of machine learning tools in Tartu.

What are they? They are computer-modeled artificial neural networks that have been made to act as similarly to the human brain as possible. Scientists in the group not only create these networks, they also make them communicate with each other and teach them to make independent decisions. “We are training these networks in hopes they will come to mimic our way of handling these processes,” Vicente explained.

Self-learning systems

The main objective is to make these artificial neural networks cooperate, not only with each other but also with humans. “For example, if in a room of two people one drops a pen, the natural reaction of the other is to help them pick it up,” the scientist said. “We want to create artificial systems that are capable of cooperating on our level and based on our nature.”

Estonian member of the workgroup Ardi Tampuu said the next major challenge is to make the artificial brains not only react to each other but also anticipate their partner’s next move. To make cooperation smoother, faster, and sporting a higher quality.

For this, systems are given self-learning capacity: they try various ways of achieving their goal until they find an optimum solution. Tampuu said that at the very beginning of the exercise all the system sees is a mass of numbers that have no meaning whatsoever. “However, once events start to unfold, and the system knows both positive and negative potential results, the agent starts to learn and spot connections between changes and consequences,” he explained.

The scientist said that agents are like children – they are rewarded for successful results that motivates them to rack their artificial brains on the problem. “We also had to rewrite a number of algorithms for a very simple reason: agents that failed to achieve their goal on the first try “took offense” and quit,” Tampuu recalled.

The team always makes sure decisions artificial systems make to achieve their goals are controlled and predictable – if only to keep them from becoming a danger to people.

These kinds of artificial systems have very similar “senses” to what we have as people – the best example perhaps is the ability to “see” or detect objects and images used in self-driving cars. Just like a self-driving car’s camera recognizes different objects its algorithms tell it to and makes the car react accordingly, agents can be put to work analyzing and processing data from the human body. Here the data can also be accessed via thousands of databases.

A practical example of this is a cooperation agreement with one of the world’s leading microscope manufacturers PerkinElmer from July the aim of which is to create intelligent software that could analyze images of human cells. Head of the UT Institute of Computer Science Jaak Vilo explained that software used to analyze images taken by microscopes has so far been based on specific algorithms, mathematical models and manmade programs. “The latest artificial intelligence technology that trains artificial neural networks using examples considerably lightens the workload of compiling new algorithms,” Vilo said.

The computational neuroscience workgroup also pursues close cooperation with face recognition software startup Markus and unmanned military equipment maker Milrem both of which employ several UT computer sciences graduates.

We must be careful

In truth, systems like these still require highly complex mathematical algorithms. Vicente explained that what is being programmed is a neural network as well as how it functions and acts. “This kind of work requires a vivid imagination, self-awareness, the ability to understand how the human brain works and redesign it. A lot of things need to be determined in the process of trial and error,” the scientist said.

Ardi Tampuu said that these artificial neural systems will not form an integral brain, while they are inspired by it. “They are still very basic compared to the brain,” Tampuu said. “Different parts of the human brain oversee different things in our organisms and are made up of very different types of cells. We are only working on a single type of neurons placed in a highly stereotypical relationship with each other.”

Tampuu agrees, however, with warnings by Elon Musk and other visionaries against uncontrollable development of artificial intelligence systems that could culminate in a situation where something starts making decisions for us. “Yes, we must be careful. That said, we are still a long way from real danger – effects of decisions made by artificial intelligences are today confined to computers,” he said.

Raul Vicente said that decisions by artificial systems are still very accurately programmed: “if this, then that”. Things will become more complicated once they are given the ability to create their own strategies.

One characteristic feature of the UT Institute of Computer Science is that it has more foreign students than it has Estonians. Ardi Tampuu said the reason for this is the fact Estonian students often quit university to start making money as developers. “They no longer even finish their bachelor’s studies,” Tampuu pointed out. “That is why we see a lot of Ukrainian and Russian students in our master’s program who are still interested in pursuing research.” Raul Vicente said that the aim of the working group he heads is to understand first and engineer second.