Papers from Centro Pi and Visgraf are accepted at NeurIPS 2025
Two papers carried out by Centro Pi (IMPA’s Center for Projects and Innovation) and Visgraf (IMPA’s Vision and Computer Graphics Laboratory) were accepted at the “Conference on Neural Information Processing Systems 2025 (NeurIPS)”, one of the most important artificial intelligence (AI) and machine learning conferences in the world. The works were produced collaboratively during Arthur Bizzi’s PhD at IMPA and combine classical mathematical methods with modern neural network techniques, with applications ranging from physics to computer graphics.
The articles “Neuro-Spectral Architectures for Causal Physics-Informed Networks” and “FLOWING: Implicit Neural Flows for Structure-Preserving Morphing”, carried out, respectively, at the IP Center and Visgraf, represent a breakthrough in the so-called hybrid approach in AI. This line of research seeks to integrate the rigor and precision of traditional mathematical methods with the flexibility and new possibilities of neural network-based models.
“The main contribution is to be able to take methods from traditional numerical mathematics into artificial intelligence, and to show that this exchange between the two fields can be very rich. AI doesn’t have to be constantly reinventing the wheel; on the contrary, this classical knowledge is one of the keys to this new technological revolution,” explains Bizzi, currently a postdoctoral researcher at the École Polytechnique Fédérale de Lausanne (EPFL), in Switzerland.
In the first paper, the researchers applied this mathematical integration to solve physics problems, especially physical wave modeling—a challenge with relevant practical implications, such as the detection of oil reservoirs through seismic wave analysis, the main application of the research.
“An important point of this paper is that we were able to use very mathematical theorems, such as the spectral theorem, which is a classic, to represent the problem in a very different way. We represent both a partial differential equation that is more difficult to solve, and many ordinary differential equations, which are easier for a neural network to solve. We used the neural network to consolidate a more promising result,” says Lucas Nissenbaum, project scientist and manager of technological projects at IMPA.
The second work focuses on image interpolation, a central problem in computer graphics. The technique presented reconciles neural networks with principles of structural conservation, resulting in more precise deformations, with potential applications in forensic medicine and even in filters used by platforms such as TikTok.
The result of a partnership between Visgraf and Centro Pi, the article was born from the convergence between Bizzi’s research on the representation of flows and the problem of morphing faces using neural networks (ifmorph). In 2024, the then doctoral student attended a lecture given by IMPA researcher and Visgraf collaborator Tiago Novello during CVPR 2024. The exhibition on IFMORP was the trigger for the development of the collaborative article that involved ten authors, eight institutions, and five countries.
“The work not only presents a new approach to the problem of image morphing, but also extends to newer 3D representations, such as 3D Gaussian Splatting,” explains Novello.
With extremely competitive acceptance rates, NeurIPS is the stage for the world’s leading advances in AI. In 2025, the event will be held between December 2nd and 7th, at the San Diego Convention Center (USA). Bizzi will participate in person, along with the Centro Pi team, to present the results of the two works.
“I think it is important to have researchers from the Global South being present at these events and ensuring our inclusion in the debate and research of this new technological revolution. We, from Latin America, are still a minority in these spaces, but the hope is that our presence will grow more and more”, points out the mathematician.