11/07/2025

IMPA stands out at leading Computer Vision event.

The article “Tuning the Frequencies: Robust Training for Sinusoidal Neural Networks” , authored by IMPA researchers Tiago Novello and Luiz Velho and master's student Diana Aldana, received a highlight distinction at the 2025 Conference on Computer Vision and Pattern Recognition (CVPR) — the leading global event in the field of computer vision. This honor is awarded to works of high relevance or exceptional quality. Only 387 articles received this distinction out of 13,008 submitted, of which 2,872 were accepted.

Held in Nashville, United States — a city nicknamed "Music City" — the conference brought together more than 10,000 participants from around the world.

The award-winning work investigates the training of sinusoidal neural networks , a type of implicit neural representation (INR) that has proven effective for modeling smooth, low-dimensional signals. While promising, these networks still face challenges in initialization and training. To address this problem, the authors proposed a theoretical framework based on a novel amplitude and phase solution of the sinusoidal perceptron. This model reveals how the network layers generate new frequencies by matching the input frequencies, allowing for robust control of the network spectrum. The proposed method, called TUNER, improves the stability and convergence of the training, resulting in planned reconstructions and avoiding overfitting.

This article was developed during Tiago Novello's time as a visiting researcher at Google DeepMind, in collaboration with André Araujo. Co-author Diana Aldana is the first master's student supervised by Novello at Visgraf (IMPA's Computer Graphics Laboratory).

"Visgraf has been dedicated to producing high-level scientific work, focusing on the main conferences in the field. Last year, we also had an article accepted at CVPR," Novello highlighted.