22/11/2024

From the Navy to mathematics, Hallison Paz completes his doctorate.

From dreaming of becoming an admiral in the Brazilian Navy to becoming a mathematician, Hallison Paz takes another step towards consolidating his academic career, marked by changes, dreams and challenges. The IMPA student and collaborator at Visgraf (IMPA's Computer Graphics Laboratory) will defend his doctoral thesis in applied mathematics next Monday (25). Under the guidance of researcher Luiz Velho, Hallison developed the research “Spectral representation of neural media in multiresolution”.

Originally from Recife (PE), Hallison Paz, 33, is a computer engineer from IME (Military Engineering Institute) and began to get closer to mathematics during his undergraduate studies. Before that, his plans were different. Hallison traded maritime routes for computer graphics and virtual interface projects.

The son of a sailor, Hallison completed his secondary education at the Naval College in Angra dos Reis, Rio de Janeiro. His goal was always to follow in his father's footsteps, whom he saw as a role model. In total, he dedicated eight years to military life, three years as a Navy officer and another five as an Army officer.

“I already had the intention of being a soldier. Not out of obligation, but my father was truly an inspiration to me. He was my professional role model. From a very young age, I thought I would be a Navy officer, a Navy commander. I always aimed high.”

The turning point to pursue new dreams came slowly. The decision to change was accompanied by a bold gamble: a career transition during Naval College. To do this, Hallison had to overcome the fear of leaving a well-established professional project and the frustrating stories of colleagues who were unable to enroll in other courses during Naval College.

Hallison was the exception. The young man was accepted into the IME (Military Engineering Institute) while attending the Naval College. He graduated as a military engineer and left the career shortly after completing his studies, holding the rank of First Lieutenant. His relationship with IMPA also began at this time, with a scientific initiation project at Visgraf after successfully presenting an academic paper at IME.

“In my second year, I did a project that was a simulator of a flying saucer that moved autonomously in space. Because of my interest in games, I wanted to work with things that had visual appeal. Mathematically, it's something more sophisticated, a simulator that solves a combinatorial optimization problem. When I got closer to IMPA, I got closer to mathematics,” Paz said.

After undergraduate research, she pursued a master's degree at the institute and now a doctorate. In this research, Paz worked on the interface between artificial intelligence and computer graphics. The highlight of her work is the use of neural networks in media representations, considering that neural networks offer a continuous representation that approximates mathematical models of signals to computational implementation.

In general, the idea of the research was to investigate the application of multiresolution theory to coordinate-based neural networks to address challenges in neural media representation, including zoom, antialiasing, and efficient data transmission.

“In computer graphics, we are interested in representing and generating virtual objects and scenes. And that's what makes up, for example, animations, cinema, and the games we know. What seemed very interesting was using neural networks as a way to represent media objects, which can be images, audio, videos, and 3D models,” Paz explained.

The result of this work is MR-Net, a family of neural networks for encoding media objects at multiple scales. The research demonstrated the potential of MR-Net in image encoding, anti-aliasing, texture mapping, and seamless material synthesis. One of the positive results was that "in image reconstruction tasks, MR-Net achieves Peak Signal-to-Noise Ratio (PSNR) values comparable to or higher than previous state-of-the-art models, using fewer parameters," the doctoral student pointed out.

One idea is to replace the models of traditional representations — in JPEG and PNG formats — with those based on neural networks. “From the way we are working, looking at the media landscape, neural networks are entering everything. There are already specialized hardware devices that can run these models very quickly. Not just train them, but mainly run them. We plan to use neural networks as a representation to, in the future, arrive at different formats. Just as there is JPEG and PNG today, there could be a format based on a neural network,” he concluded.

Coming from a humble background, his doctoral research is seen as yet another achievement attained through much effort in the young man's career. In addition to this, Hallison has already accumulated various opportunities, such as an internship at Meta . He is also the founder of the successful YouTube channel "Programação Dinâmica" (Dynamic Programming), with over 200,000 subscribers, where he produces content about technology, focusing on programming, data science, machine learning, and artificial intelligence.

Now, Hallison is planning new paths for his post-defense career, which include a trip to Japan to participate in Siggraph Asia , the most important computer graphics conference. Currently, there are two editions per year, one in North America, where Hallison presented a paper in July 2024, and another in Asia. Furthermore, the young man affirms that he will continue to nurture his professional dreams and his passion for knowledge. "If I could have all the knowledge in the world organized, I would do it."