Gonzalo de Polavieja, neuroscientist: ‘We tend to follow the few who make clear decisions’

Gonzalo de Polavieja, 56, is exasperated by the ease with which many people opine on topics without knowing anything about them. A neuroscientist trained at Oxford and Cambridge, with a PhD in quantum physics and a postdoctoral degree in mathematical neurobiology, he is currently on leave from Spain’s CSIC research center and directs the Laboratory of Mathematics of Behavior and Intelligence at the Champalimaud Foundation in Lisbon, where he studies how groups of animals — including humans — organize themselves. The scientist speaks cautiously, trying to avoid mistakes with the same discipline he applies to studying neural circuits. After the interview, he wonders whether if his answers did justice to questions on topics he knows less about.

In a recent lecture at the summer courses of the Menéndez Pelayo International University (UIMP) in Santander, titled Applications of AI in Neuroscience and Behavior, he presented his work with the zebrafish (Danio rerio), an animal model increasingly used in biomedicine and neuroscience. Its transparency at early stages allows researchers to observe the brain in action; its nervous system — with about 100,000 neurons — is simple enough to study as a whole, and it shares more than 80% of its genes with humans. All of this makes it a privileged tool for understanding how neuronal activity translates into observable behaviors such as learning, exploration, or decision-making.

Question. What excited you about the interaction between neuroscience and artificial intelligence (AI)?

Answer. My career began with quantum chemistry, but I’ve always been drawn to biology. Since I’m not good at memorizing, I became interested in mathematics, which allows me to compress information. During my postdoc at Cambridge, I was able to make the leap into neuroscience, and since then, I’ve been trying to understand it with simple mathematical models.

The problem is that in biology, problems are very complex, and that’s where AI is useful, although it carries the risk of giving you solutions without you understanding them. For me, the key is for science not only to solve a specific problem, but also to allow us to think about the next experiments. That’s why we use AI in tasks where understanding isn’t essential, such as animal tracking, but when it comes to understanding processes, we need more transparent mathematics. I’ve been thinking for years about how to achieve this.

Q. Isn’t it paradoxical to try to unravel the intricacies of the brain using an AI that we don’t fully understand either?

A. Conventional AI predicts, but it doesn’t explain. To help us understand, it needs to be simplified. In my case, I transform neural networks [digital algorithms that attempt to model the functioning of biological neurons] into more interpretable modules and check that they continue to predict in the same way. Modeling means understanding, and that requires abstractions: we’re not looking for a copy of the brain, but rather representations that allow us to think about it.

This simplicity is strategic, not real, but it is the only way to advance in the understanding of such complex systems.

Q. As a study model, does the zebrafish help with that simplification?

A. Yes, it has many advantages. In the larval state, their brain is transparent, and we can record the activity of the entire brain without opening the animal, under normal physiological conditions, which is not possible in almost any other species. Furthermore, their genome is sequenced, which allows us to design sensors to visualize neuronal activity. All of this allows us to study their behavior very well. I previously worked with invertebrates like the Drosophila fly, but I always look for the simplest model to understand a problem. Understanding the brain is very difficult, and the only way is to simplify it as much as possible.

Q. What have you learned about how zebrafish behave in groups?

A. In very simple situations, fish tend to choose to move towards where there are more individuals, but in natural conditions that almost never happens; the behavior is much more complex. When analyzing it with models and tracking techniques, we discovered that they don’t necessarily follow the majority, but rather a few individuals (usually one, two, or three) that are ahead and move with more conviction or speed. In other words, the group is usually guided by a small, very determined subgroup. To reach that conclusion, I had to rely on AI designed to be interpretable, so that it not only predicted the behavior, but also helped me understand the rules behind it.

Q. Can these models be applied to group behavior in mammals, such as humans?

A. Yes. Although we conducted the experiments with zebrafish, we designed general models, not specific to that species, and we found that they also work in ants and humans. In comparable situations (for example, when a group must choose between simple options without language involved), the same rules apply: individuals tend to follow the few who make clear decisions. The beauty of these simple models is that, when they work, they are usually valid across many species.

Q. The fish follow the leader?

A. Yes, but it’s not that simple. The “leader” is the one who shows the most conviction, moving quickly, even if it may be for the wrong reasons, and the others follow because that strategy has worked for them from an evolutionary standpoint.

Q. Do you mean that this individual, zebrafish or human, is perceived as reliable in its decisions?

A. Exactly. There’s a parameter in the model that precisely reflects the reliability of who you follow.

Q. What promising applications do you see in AI for neuroscience today, and what principles of neuroscience could, in turn, improve AI algorithms?

A. A very interesting line of research is the so-called foundational models, built from brain activity data of a species (mouse, zebrafish, etc.) in multiple situations. They function as a common foundation that other researchers can adapt, even as a partial substitute for experiments. Regarding the reverse influence, many neuroscience findings could improve AI, but the transfer is slow: verifying how the nervous system actually works takes a long time, while AI is advancing at a dizzying pace. Even so, this exchange has already borne fruit and is likely to continue to do so in the future.

Q. What potential does quantum computing have in the evolution of neuroscience and AI?

A. As far as I know, quantum computers are only superior in a very narrow set of algorithms; they don’t automatically do everything classical computing already solves better or faster. So it’s not guaranteed that AI will be of benefit across the board. We’ll only see a difference if someone finds an algorithm where quantum computing offers a real advantage.

Q. What are you particularly interested in regarding AI research?

A. Together with Fernando Martín Maroto, another researcher on my team, we are developing a completely different approach to neural networks, based on abstract algebra. The idea is to create an AI that learns powerfully but, unlike current models, is transparent and understandable at every step. It is not based on minimizing errors as neural networks do, but on algebraic properties that guarantee convergence to the underlying rule in the data. The big challenge now is to transform this promising mathematical foundation into a practical system that is as efficient as current models and, at the same time, much more interpretable for humans.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition