Understanding the human mind without a human mind : The AI neuroscientist | Romy Lorenz | TEDxNTUA

Translator: Romy Carvalho Reviewer: Peter van de Ven How does matter give rise to the mind? This is probably one of the most fundamental questions in philosophy and one of the earliest questions of mankind. As early as a few centuries BC, Plato and Aristotle in the West and the Buddha and Laozi in the East were contemplating on the workings of the mind.

Out of the human thirst for knowledge about their own mind, the scientific discipline "psychology" emerged. And quite suitable with the location of the event today in Athens, the word has its origins in Greek and means "the study of the psyche. " In the past, our measurement instruments to look into the human psyche were limited.

We had to take rather indirect routes - for example, by observing human behavior, measuring how accurate they perform a certain task. However, over recent decades, developments in human brain imaging technology have helped us to gain remarkable direct insights into the living, human brain. And with these developments, a new scientific discipline formed: cognitive neuroscience - bridging neuroscience and psychology.

I'm a cognitive neuroscientist, and even after thousands of years, this discipline is still interested in the same questions that preoccupied Plato and Aristotle. How does the brain give rise to the mind? Or more specific, How do networks in the brain give rise to the mental abilities that made us humans the successful species we are today?

Abilities such as producing language, solving problems, fluidly adapting in the face of change, learning, or our remarkable capacity to think about thinking. But how do we cognitive neuroscientists approach such a big question? And the answer is less glamourous than the question itself.

And as you can probably fortell, scientists are reductionists, and we like to take complex phenomena and break them down into smaller parts we can study in the lab doing experiments. So, how do the experiments look like ? Usually we start with a research idea.

Let's say we're interested in understanding which regions in the brain are involved in humans behaving intelligently. So, we will choose a task that measures an aspect of intelligence, for example this one: here the challenge would be - that is like a classic, logic task - and the challenge would be to find the set of shapes that is different from the other ones. Now, after this, we're actually conducting the experiment; that is we're putting our subjects, our human participants, into the brain scanner, in which they're performing this task.

And now, while they're lying there, every two seconds, we're taking a 3D snapshot of their entire brain. And every one of these 3D snapshots consists of over 100,000 data points, measuring brain activity at that moment in time. And after all of this, after we collected all the data of all the participants, the real fun starts.

We have to sift through this massive amount of data, trying to detect tiny changes in brain activity, using sophisticated statistical analysis. And in the end, we're getting these wonderful, colorful images you've probably seen in the media before. And this image now tells us which regions in the brain are associated when humans apply logic to a problem.

So far, so good. But what if I tell you understanding the human mind is not as easy as I'm advocating here. And in fact, other experiments showed that the same set of regions was also active when we're memorizing a telephone number, dividing our attention between driving and the navigation system, or calculating our share on the restaurant bill.

So you've seen, after a decade of trying to understand the set of brain regions, we haven't learned that much except it's involved in almost all daily activities. As you can imagine, that's rather unsatisfying for a cognitive neuroscientist. We're using high-tech brain scanners in order to learn detailed specifics about this network, and what we're left with is a rather unspecific and confusing picture about the human brain.

The problem with this lies in the way we're currently conducting our experiments. Remember, I told you, commonly, we just use one task we have our human lab rats performing in the scanner. So, after the experiment, we can only draw conclusions about this one mental ability this task is measuring.

However, in order to understand the true role of this network in intelligent behavior, we have to look at many mental abilities that all measure an aspect of intelligence. Only by directly comparing these mental abilities, we're able to say this network is more active when we apply logic to a problem, compared to memorizing, compared to dividing our attention. Let me give you a real-world analogy.

Let's assume this network of regions is a new couch we just bought, and the mental ability we want to test in a single experiment is the color of the walls in our living room; and our aim is we want to match this color in the living room with our new couch. So what we're currently doing in the field is we're selecting one single color, we're painting all the walls in the living room with this single color, put the couch back in, and then we're like, "Mmh, maybe there's another color out there that is better. " So, we're running a new experiment, choosing a new color, single color, and again paint all the walls in the living room with this color.

But our intuition would tell us, "Wouldn't it be much more clever, smarter, and efficient if we bring color samples at home, compare them with each other and having the couch next to it? " But our technology limits us here. Unlike color samples, brain scanning is expensive and slow.

On top of this, humans fall asleep rather quickly when they have to lie completely still in a dark room with a steady buzzing sound in the background. So these limitations force us to only really test a handful of mental abilities at the same time in a single experiment. And these limitations give rise to a rather confusing picture we have about the brain right now.

But in addition to these technological limitations, there's another very serious source of confusion in our field. And this source of confusion is us, the scientists. Or to be more precise, the human scientists.

We human scientists are not as objective as we're trying to think. We're fooling ourselves, unconsciously, with the same mental biases that we're trying to investigate in our experiments. So let me give you two examples.

Here we have the IKEA effect. This refers to people in general placing artificial high value on products they've built themselves, like a BILLY shelf. But this also applies to science.

Scientists trust more the analysis pipelines they coded over months or the logic task they designed themselves. Even though there might be a bug in it and it would produce faulty results. Then, we have the Texas Sharpshooter effect.

The name goes back to a joke of a Texan firing gunshots at the side of a barn, then he goes there and draws a center, he draws a target centered around the tightest cluster of hits, and then he claims he's a sharp shooter. And this refers to scientists having massive amounts of data sets but only focusing on a subset of the data that looks like a statistically meaningful pattern and ignoring all the random noise around it. However, these mental biases are as old as science.

What has changed, though, is that the academic environment has become more and more competitive. In order to survive in academia, researchers have to pile up applications with statistically meaningful results. In addition to this, we're getting bigger and bigger data sets.

For example, with brain imaging. And there's not just one possible way to analyze this brain data. There are hundreds of decisions we have to take, and they're all equally valid.

And given the computational resources available today, we can just run all of them, easy, quick and cheap. But the combination of mental biases and academic pressure is explosive. And it has resulted in many questionable research practices where researchers are repeating their statistical analysis in subtle, different ways until they find a result that looks statistically meaningful and thereby is publishable.

So they can succeed. So - Yeah, en masse and over many years, this has resulted in our field being flooded with published findings that are actually not true. Nobody can replicate them.

We refer to this as the reproducibility crisis in science. And it's important to note that it's not just cognitive neuroscience that is affected by this, but actually many scientific disciplines. According to a survey, 70 percent of researchers tried and failed to replicate another scientist's experiment.

And the growing recognition that seemingly well-established results failed to replicate has shaken many scientific disciplines at their core. Because reproducibility is one of the pillars the scientific method is based on. But don't worry, science is not all doomed.

There are now many great initiatives that try to fight the reproducibility crisis; and they will help to clean up science. They all have in common, though, that they're trying to keep the human error under control by making changes to the system and the infrastructure of publishing. And that's exactly where my research comes in.

I and my colleagues from Imperial College London, we took a different stance on the problem. We looked at the root of this problem. We asked ourselves, "Why not removing the human neuroscientists altogether?

" And that's how we came up with the AI neuroscientist. And the key difference between the human neuroscientist and the artificial intelligence one is that the AI does not wait to analyze the brain data until it has been collected, like we did before; instead, it analyzes the brain data in real time, on the fly, while the subject, the human participant, is still lying in the brain scanner. And this comes with a serious advantage as it makes brain scanning much more powerful and flexible.

Now we can test many mental abilities at the same time. And to explain you why, let's go back to our couch analogy. So, the AI is doing exactly what our intuition would've told us.

It starts by testing very different mental abilities, such as logical thinking and dividing our attention. This would correspond to colors at the opposite end of the color spectrum, like blue and yellow. It can then quickly decide, in real time, which one of these mental abilities is more heavily engaging this brain network we're interested in or which color fits best.

Let's say logical thinking or the color blue has done this. So now the AI goes on and only tests mental abilities that are very close to logical thinking, such as calculating. Going back to the color spectrum, it would only test colors that are closer to the blue, such as green, or different shades of blue.

By this, you realize the AI can really efficiently and quickly learn which colors are best for our living room situation, or which mental ability best describes this brain network. And the trick here is, the AI does not need to test every single color out there. But it leverages knowledge it has about the similarity of colors and their space in the spectrum.

I'll illustrate you with this little video. So, here the AI is embodied as a little robot; it's not an actual robot. And now the AI starts by selecting one out of many possible mental tasks, or mental abilities it will test.

So here, bottom-left corner. It then will analyze - it will first measure and then analyze brain activity in real time. And then, in this case, it will see, "Okay, this mental ability did not activate this network very strongly," illustrated by the blue color now, in the next slide here.

So, now the AI will understand that it can probably dismiss mental abilities similar to this one and move on to test a very distinct one, far away from the one we just tested. Again, it shows that this task to the subject in the scanner measures and analyzes activity in the brain. And here it sees, "Wow!

This was much better! We're getting much higher activation in this network. " So, now it will only test mental abilities that are very close to this one.

And with this and over time we will learn very quickly which mental ability best describes this brain network. And the efficiency of the AI doing this comes from the fact that it analyzes the brain data in real time. And this has another wonderful side effect.

As we humans can't do what we usually do anymore, which is waiting until it has been collected and then torturing the data until we get the results we want to publish. No, the AI is running the experiment. So, we humans are just passively watching, our hands are tied, we can't cheat, as we've been completely taken out of the experiment loop.

And we hope that the AI neuroscientist will help to accelerate our understanding about the human mind without being affected by the biases of the human mind. And the AI neuroscientist differs in one core element from many other AIs you've probably heard in the media before. It is not trying to mimic human behavior.

It's built to go around pitfalls of the human brain. And this is also my take-home message for you today. Let's dream about a future where we build AIs that are better versions of ourselves.

That are not fooled by gender or racial biases, or by irrational fears, or even the illusion of an ego, as the Buddha had already understood. Thank you.