Mathematics to "make the population trust more in artificial intelligence"

When Paula Gordaliza began working on the application of statistical theories to create fairer algorithms, in 2017, artificial intelligence (AI) was still being talked about in the future tense.. Today, scientific articles on the topic accumulate every week and debates about the increasingly central role that algorithms occupy in society are omnipresent.

His line of research is at the heart of one of the most important conversations: his work aims to design and analyze machine learning methods that detect, control and correct possible biases in the results, which will contribute to creating fairer algorithms, eliminate discrimination in the results and, ultimately, “make the population trust more in artificial intelligence”, in his own words.

Trained at the universities of Valladolid (UVa) and Toulouse III-Paul Sabatier, she is a postdoctoral researcher at the Basque Center for Applied Mathematics (BCAM), in Bilbao, and an associate professor at the Public University of Navarra, positions from which she claims the role fundamental of mathematics. “Many professionals are involved in artificial intelligence, but the basis provided by mathematics is what allows us to solve problems,” he says.. This year, his work has earned him recognition from the Vicent Caselles Mathematical Research Awards granted by the Royal Spanish Mathematical Society and the BBVA Foundation.

We tend to think of machine results as something neutral, objective.. What does it mean to talk about biases in artificial intelligence? In reality, artificial intelligence does what it is told to do, nothing more. But to work it needs data and some of the main biases that appear in the results actually come from the databases with which it works, which are often the ones that are biased, for different reasons.. In recent decades, an effort has been made to take care of the data collection process, but many times we find certain groups that are more represented than others and that conditions results. So the biases are in society, in the human part of the process. We are seeing it in some of the ChatGPT results, which privilege English over other languages or which can perpetuate certain stereotypes in genders or social groups. In part yes, because the algorithm that powers the AI learns from the data we give it and Any bias that that information contains will also be learned (and in some cases, even worse).. Those examples that you mention are one of the ways in which it manifests itself, because it is common in the new artificial intelligences, called generative, where it is seen that it is necessary to do work to detect and reduce these biases. And what role does statistics, and in particular the theory of optimal transport, in that effort? Like many other answers in mathematics, the theory of optimal transport arose to answer a real and concrete problem. In this case it appeared with the so-called Monge problem, to know how to transport a volume of mass in the most effective way possible (from a pile of sand to a hole).. As a result of this approach, an entire theory is developed and applied to the field of statistics, where it is very important in relation to the concept of probability measurement: In short, it is about transferring one probability measure to another with the least possible effort. How is it related to AI? Regarding AI, in the simplest case we can have a database with a sensitive variable, such as gender. And we want to erase the influence of that variable, so that when the data is used the algorithm cannot reproduce biases.. So what we try is to match the rest of the characteristics. Specifically, our idea is to transform the two original distributions of men and women, making them as similar as possible in the rest of the variables and thus erase the possible influence of that protected attribute. The fact that many of the AI tools that are based on the machine learning publish results without being able to know the process (the so-called black box problem) is one of the sources of distrust towards this technology. That is the great objective of mathematics, trying to apply rigor through theory, obtaining results powerful that open those black boxes and that society can trust more in the process. Sometimes these solutions that algorithms give us can be attractive, because it is seen that they work, but they lack sufficient traceability to know how they are reached. It is a field where public-private collaboration is fundamental, what is investigated has a clear and immediate practical application for society. How do you see this company-university duality? Research is always ahead of applications, especially in mathematics. But it is true that there is an increasing demand from companies for methodology, strong methods to solve problems in society, which are increasingly complex and multidisciplinary and which also require multidisciplinary teams (not only mathematicians and statistics).. I think it is positive that the industry offers us more and more real problems, which serve to motivate research, because this type of synergies is something that is of increasing interest. And what is missing for these synergies to increase? I think they come together different issues: on the one hand, companies are very protective of their own data or those of their clients, which can generate problems with confidentiality. There is also the fact that companies have their own methods and processes, for which in some cases they pay a lot of money and which are difficult for them to separate themselves from. Mathematics in particular is a discipline that is difficult to make visible, despite its importance. Mathematics has had a bad reputation since school, since we were little, because it is considered difficult. But the truth is that they serve to solve problems and today we find increasingly complicated problems.. We must try to get the message across that mathematics is not only necessary, but is useful for everything.. They are a constant that is behind any advance and any science.. That is why I believe that many times young people who are considering what profession to choose, what to train in, feel attracted to other disciplines that are more applied than basic mathematics.. In my point of view, specialization is forced too soon.. Skills and basic knowledge should be deepened. So that young boys and girls who are attracted to artificial intelligence, for example, have a foundation in mathematics, statistics and other areas. And when they are ready, with a toolkit of those basic concepts, they can make the leap to more specific concepts.