**IMFD present at international workshop “When Deep Learning Meets Logic”**

April 2022.- Pablo Barceló, deputy director of the Millennium Institute Foundational Research on Data (IMFD), researcher at the Centro Nacional de Inteligencia Artificial (Cenia) and director of IMC UC, was invited to the international workshop “When deep learning meets logic”, a virtual event sponsored by Samsung to be held in June. In its first edition, held in 2021, this event brought together more than 1,000 participants from prestigious institutions such as MIT, Google, IBM, Microsoft, and the universities of Oxford and Cambridge.

Leslie Valiant is not only a professor at the prestigious Harvard University (USA), but also a recognized pioneer in the study of computation theory and, in 2010, received the Turing Award, a distinction known as the “Nobel Prize of Computing.” Christos Papadimitriou, a professor at Columbia University (USA), also holds several awards such as the Gödel Prize, the Knuth Prize, and the John Von Neumann Medal, in addition to being the author of the book Computational Complexity, one of the most widely used and referenced works in the field of computational complexity theory. These two internationally renowned researchers were in charge of delivering the opening and closing talks, respectively, of the first edition of the workshop “When deep learning meets logic,” held last year, and whose 2022 edition will feature a presentation by Pablo Barceló, director of the Institute for Mathematical and Computational Engineering of the Universidad Católica (IMC), who was invited to participate as a speaker.

The workshop, which will be held virtually from June 6 to 8 and is sponsored by Samsung Research, is organized by Vaishak Belle – a researcher at the University of Edinburgh (Scotland) whose studies focus on the intersection between machine learning and logic-, Martin Grohe -a mathematician at RWTH University (Germany) and an expert in fields such as database theory and finite model theory- and Efi Tsamoura, a researcher at Samsung AI whose studies focus on topics such as distributed databases. The first edition of the event drew more than 1,000 participants, a turnout that, according to the organizers, reveals the high level of interest surrounding the efforts made in recent years to integrate the fields of logic and deep learning.

In this regard, Barceló -who holds a PhD in computer science- explains that artificial intelligence began to develop about 50 years ago along two major paths. “One is the connectionist approach and the other relates more to knowledge representation. In the first, one could say, speaking in fairly general terms, that what we are going to do is replicate a brain. A brain has neurons that connect, but we’re not going to explain how to do it. We’re going to build, in a certain way, the infrastructure and then, based on the data we have, we’ll make those neurons connect and learn something. When new data arrives, those neurons connect in a different way and learn something else.”

That approach, adds the director of IMC, is the one that has given rise to crucial developments such as deep learning, very deep, multi-layered neural networks that solve highly specific and complex problems much more efficiently than a human being. “Typical examples in this area are image classification and text processing, such as that offered by Google Translate, which works almost perfectly.”

Alongside the so-called connectionist approach, another path developed that for a long time was even more predominant, one related to knowledge representation. “That is, how do I explain to a machine how to solve certain tasks. That has to do with how to represent the knowledge or domain I have about that task, and that specification is commonly given in a logical language,” says Barceló. In the case of artificial intelligence, logic is also a language for explanation: “Neural networks are often built, but we don’t fully understand what happens inside them. It’s like a black box. But logic can give us a concise representation, an explanation of what it is doing.”

Although the connectionist line has seen enormous progress, there are also signs that it is reaching its limit, since it still fails to understand context. “It’s very inefficient. It needs trillions upon trillions of data points to learn something that may not even be that complex, because it reasons very little and provides very few explanations. On the other hand, we have this other logical line that explains a great deal, but at the same time is extremely costly and inefficient from a computational standpoint,” explains Barceló. For that reason, the researcher adds, it is now believed that the only way to achieve a new breakthrough in artificial intelligence is by combining these two approaches, so that “deep neural networks understand more about the domain they are learning about, so that they can learn with less data, while also providing explanations.”

That meeting point is what is known today as Neuro-symbolic AI, where, as Barceló points out, logic plays a “predominant or particularly important role, because it is the language we use to represent knowledge and obtain explanations.” Indeed, the organizers of the virtual event state that this is the first workshop to cover all aspects related to the integration of deep neural networks with logic and symbolic computation.

Furthermore, according to a statement Efi Tsamoura prepared for the first edition of the workshop, this event not only covers a wide range of topics related to this subject, but is also pioneering in presenting both foundational and applied research in the field. “Although the integration of these two paradigms has been the subject of various workshops associated with organizations such as AAAI, IJCAI, NeurIPS, and CVPR, ours is the first that does not focus on a single topic but instead offers a more holistic view of this fascinating research area that combines deep neural networks with symbolic systems,” she said.