As soon as I achieve my goal, I will connect my brain to external interfaces. I am most interested in what would happen if I connected to another person and to their brain, says Andrzej Banburski, PhD, from MIT in conversation with Monika Redzisz
Monika Redzisz: Although you’re a scientist, you ask philosophical questions. For example: ” One of my biggest motivations is to determine if there is a scientific framework to answer the question ‘Why is there something rather than nothing?'” as you write on your website. Can science answer such questions?
Andrzej Banburski*: I’ve always been asking myself philosophical questions. Why is there something rather than nothing? How does the universe work? This is why I got interested in physics already in high school when I was still a student in Gdynia. I came to the conclusion that I would come closest to answering these questions if I studied the theory of quantum gravity, which is a field combining quantum mechanics and Einstein’s theory of gravity. It’s the Holy Grail of physics. This theory does not actually exist, and yet we know that it has to exist, because there are at least two points in the universe in which it has to apply: black holes and the Big Bang. My PhD thesis focused on loop gravity, which is one of the theories of quantum gravity, which, as a matter of fact, does not exist either. We have models that only roughly interact with what we can observe in astrophysics. As long as we are unable to come up with a solid theory, we can’t calculate anything.
Calculate what, for example?
For instance, let’s say that we have a hadron collider on a galactic scale and want to create a black hole. We have no idea what could happen. We can’t calculate if it is possible to travel in time or to travel faster than light. If we had this theory, we could answer a lot of very interesting questions. And this is why I decided to create it. But the truth is that it is an extremely difficult task …
In what way is it connected to what you are doing now? You work in the Massachusetts Institute of Technology and concentrate your efforts on artificial intelligence. Why did you get interested in it?
I always thought that physics would be my thing, but during my PhD studies at the Institute for Theoretical Physics in Canada I had an opportunity to try something different. I was invited to go on holidays to Microsoft. They needed people who would develop applications for their new product called HoloLens. I was intrigued by that. I had several ideas but I didn’t know exactly what that technology could actually do. I began to wonder what it would be like if, having access to three dimensions, we did not have to perform mathematical calculations on paper or on a blackboard. Would we perceive mathematics differently?
Did you hope that you could transfer math to 3D? But why would anyone want to visualize math equations?
The visual cortex is the biggest part of the cerebral cortex. I thought that if we could transfer mathematical operations to three dimensions, we would think faster and solve math puzzles more efficiently.
If we could transfer mathematical operations to three dimensions, we would think faster and solve math puzzles more efficiently.
Can you even locate mathematical thinking in the visual cortex?
We don’t know which areas of the brain are responsible for dealing with numbers, but mathematics is not limited to arithmetic. I have always been convinced that mathematicians imagine equations whenever they are involved in crunching numbers. I think that if they had the opportunity to visualize them and if they could enter a space where such objects were visible, they would discover more possibilities. I wanted to make a platform where two mathematicians could work together, which is rare, because mathematics is difficult to talk about. So I developed HoloMath application, which allows to visualize data. When you put your glasses on and launch the application, a three-dimensional menu appears. Everything is voice-controlled. If several people wear HoloLens glasses, they can all see the same thing. You can walk among objects and perform various operations.
Can you see and perform any mathematical operation?
Yes, although you are capped by a certain level of complexity. For example, categories are difficult to visualize, but data, functions, and matrices are quite easy to handle. You can look at them from different angles and think spatially.
Can you use this application?
Unfortunately you can’t. I am the patent owner, but you can’t buy the app. Frankly speaking, it’s my fault. I should’ve published the results. I’m sure that if I had kept the ball rolling for a bit longer, it would’ve been a really useful tool for many physicists and mathematicians … But my plan was only partially successful. It’s hard to talk about math without writing. My application might be palatable to mathematicians, provided they use the same symbols, but others might have a problem with it.
I also had to find a system that would be able to use symbols. There was one I needed on the market – programming language for mathematics. We got their license but they were hard to do business with. Besides, I had to go back to Canada to my institute. Our director even wanted to establish a separate department that would take care of it, but eventually the idea fell through due to financial reasons.
You had just taken your PhD. What was next?
I asked myself what interested me most. I came to the conclusion that it was still theoretical physics. However, I felt that I, as Homo sapiens, was not very well adapted to solve such problems. We are evolutionary unique. We are the only ones that can do mathematical research. Let’s take a look at a galactic hadron collider, for example: it would probably take me about 10,000 years to experimentally verify a possible theory of quantum gravity! I didn’t want to work on something that I won’t live to see. I began to wonder what it would be like if we were more intelligent. Or immortal. How can you get there? Maybe if I could connect to a computer …
I started reading about the interfaces between brains and computers. We are already able to control robotic prostheses. What if we could control an external artificial brain? The artificial brain has no limits when it comes to size, it doesn’t have to fit in the skull, I don’t have to carry it with me. I could have an implant placed in my body and connect to it via the internet. Maybe at some point I would use it more than my own. Theoretically speaking, I could transfer everything that is in my brain to it. Isn’t that being immortal? I decided that I had to build it. That’s how I came to the idea of artificial intelligence; and then I received an offer from MIT. In fact, I wanted to work here ever since I could remember; when I was a kid I used to play a computer game called Half-Life in which the main hero was an MIT physicist. In 1998 it was the game of the year.
And on top of that MIT is one of the best places to conduct research on artificial intelligence.
Exactly. I was invited by professor Tomaso Poggio, who wanted to understand how deep neural networks actually work. They work pretty well, but we don’t know how. Mathematically speaking, we cannot explain it. Our understanding of that phenomenon can be compared to what we know about neurons in the brain. We know that they have layers, that they do something that resembles data filtering – much like in the first layer of the visual cortex, in which neurons perceive oblique lines, and subsequent layers respond to increasingly complex patterns.
We are already able to control robotic prostheses. What if we could control an external artificial brain? Maybe at some point I would use it more than my own
It is said that neural networks are black boxes.
Precisely! Typical networks are characterized by millions of parameters. Although they are statistical models, they are out of line with the basic principle of statistical models which is taught to first year students of statistics and which says that there has to be fewer parameters than data. In artificial neural networks we have roughly 50,000 pieces of data and millions of parameters. The more parameters we have, the better the networks work. Why? We don’t know. This is similar to what is happening in our brain. Each neuron/parameter is not important in itself. They are similar, but the more we have them, the more stable the network will be.
Can you open a black box?
I’m working on it. I’m trying to create a network that would explain why it generated this and not another result. When we show it a picture of a cat, it will give us reasons why it classifies the picture as a picture of a cat. „This is a cat because it has a tail, fur, ears etc.” When it makes a decision, it’ll be able to give a reason for it. “You will not be granted a loan because this and that”. If the reason makes sense, that will be enough.
If that happens, can they be trusted?
Yes, but to a limited extent. The problem with these networks is that they can be easily hacked. Let’s say we’re talking about a neural network responsible for driving a car or for classifying pictures. A small modification, completely impossible to notice by humans, can change the way the network works. We show it a cat and it tells us it is a plane.
That sounds dangerous.
Very dangerous. At MIT we printed a small toy turtle which was recognized by a slightly modified neural network as a rifle. We can also do it the other way round – an airport’s neural network designed to ensure security can take a rifle for a children’s toy and confirm that the object it has just scanned is safe. An attack of that sort has already been performed against Tesla. A small sticker stuck on a STOP road sign resulted in car failing to correctly recognize the sign. It is fairly easy to cause such interference. Those networks are extremely unstable.
Can you make them more immune to attacks?
We are working on it.
The blame is often on bad biased data; you are talking about an internal weakness – instability.
Yes, because I have reasons to believe that data will always be a problem. No one is fully impartial. People who create these networks are biased too, even if they are not fully aware of that. There is always something that you love and something that you hate. If that wasn’t the case, we wouldn’t be able to make any decision. But that means that systems will never be neutral.
Is it even possible to avoid this instability? The more complicated something is, the more exposed it is to any sort of interference.
That’s true but humans are not so easily deceived. It is not simple to convince us that we see something different from what we see, even though our visual cortex is quite complicated. We started to think about it. What does the human brain have that models don’t? Let me illustrate that with an example: because of the fact that the highest resolution is in the center of our field of view, we need to look at an object several times to understand what we see. Neural networks don’t work like that. We need to find more of such elements.
Unlike deep neural networks, people can learn from very few examples. So maybe our brain is not a tabula rasa at birth? After a child is born he or she does not see much, but can allegedly recognize the pattern of a human face.
Have we been programmed by the evolution to recognize human faces? Interesting studies have been conducted on monkeys which were isolated from others right after birth. People who looked after them had their faces covered, so the monkeys did not see any face for the first few months. When they finally rejoined the herd and their guardians removed their masks, the area of the brain that activates in the case of small children at the sight of a face did not activate in the monkeys’ brains. The question is whether it got activated at the sight of something else. And the answer is yes. It did activate at the sight of hands. They associated hands with something that fed them and protected them.
I’m trying to create artificial neural networks that will be able to reason in a symbolical, abstract and logical manner
The experiment has proved that it is not an innate reflex. Children learn it in the first days of their life because they associate their guardian’s face with food and safety. The truth is we have to learn almost everything. The thing is that up to the age of three we get a lot of information and billions of images and pieces of data.
I remember you wanted to improve our intellectual capabilities. Was there any particular purpose to do that?
To be able to solve mathematical problems faster. I began to wonder if it was possible to create an artificial mathematician – artificial intelligence which would be capable of solving mathematical problems. So far, neural networks have been terrible at symbolic and abstract reasoning. They can’t even add up numbers without making a mistake. Well, maybe they can add within the range from -400 to 400 but can’t go beyond. And that’s a huge problem. Our brain has two reasoning systems that work together: the first one might be called associative, the other being more logical. Neural networks are based only on associations and are not able to perform any logical operations. No neural network is able to deal with any mathematical proof. That’s my second subject at MIT: I’m trying to create artificial neural networks that will be able to reason in a symbolical, abstract and logical manner. I think I will work on that concept for the next couple of years.
I wonder what will be next.
As soon as I achieve my goal, I will connect my brain to external interfaces, although I’m not entirely sure if I will succeed in creating a sufficiently compatible architecture. In fact, I am most interested in what would happen if I connected to another person and to their brain.
Would it be possible for us to think together? Would it be possible for me to see his or her thoughts? I assume that if we created a network of people connected to one another, we would be able to solve problems that are more complicated than the ones we are struggling with today. Will that sort of “internal” cooperation be possible? We sure do work together but there is still the language barrier, which slows us down. We think faster than we can communicate. Maybe we would be able to answer questions that we cannot answer today?
For example how to deal with global warming.
That’s a great idea.
I know. I always have great ideas. But all this is really happening. You start with science-fiction and after some time science-fiction becomes the reality.
*Andrzej Banburski, PhD works in the Center for Brains, Minds and Machines and in the Poggio Lab at MIT. He studied mathematical physics at the University of Edinburgh and in the Perimeter Institute for Theoretical Physics in Canada, where from he took his PhD in quantum gravity.