Toss Paper on AI Model Training Optimization Accepted at Top Global Conference

• Presents large-scale federated learning technology that optimizes AI model training in environments where personal information protection regulations prevent data from being centralized on a server
• The research shows privacy and model performance can be achieved simultaneously

Toss announced today that a research paper authored by one of its employees has been accepted to the Neural Information Processing Systems (NeurIPS), one of the world’s most prestigious academic conferences in artificial intelligence. The study was conducted by Jinu Lee, Machine Learning Engineer of the Toss Face Modeling Team, in collaboration with Seoul National University’s Computer Vision Lab.

NeurIPS is widely regarded as the most influential conference in machine learning and neural information processing, with a paper acceptance rate of only around 20 percent. This year’s conference is taking place from December 2 to 7 at the San Diego Convention Center in the U.S., bringing together researchers from around the world to share the latest AI advancements.

The accepted paper titled “FedLPA: Local Prior Alignment for Heterogeneous Federated Generalized Category Discovery” introduces Federated Local Prior Alignment (FedLPA), a Federated Learning technology that optimizes AI model training in environments where personal information protection regulations prevent data from being centralized on a server. It addresses a key limitation of conventional federated methods, whose performance degrades significantly when data characteristics vary widely across clients^* or user groups, or when encountering entirely new data types.

The research team combined Infomap-based local clustering—which automatically groups data with similar characteristics across clients^* or user segments—with a Local Prior Alignment technique that stabilizes training through prediction alignment. This approach allows each device to autonomously understand and leverage its own data distributions, demonstrating strong performance in Generalized Category Discovery (GCD) by accurately identifying novel categories even when category types or data distributions are unknown in advance.

The study further shows that privacy and model performance can be achieved simultaneously: FedLPA enables the development of global AI models that comply with the legal requirements of countries with strict data-protection regulations.

“This research focuses on optimizing federated learning to remain effective even in highly constrained environments where data cannot be moved to a server due to regulatory constraints, each client^* has a different data distribution, and the number of new categories is completely unknown,” said Jinu Lee, Machine Learning Engineer at Toss.

A Toss spokesperson added, “It is a significant milestone that Toss’s AI capabilities have been formally recognized for the first time by a world-class academic conference. We will continue advancing practical, service-ready technologies to deliver more sophisticated AI-powered services while upholding strong privacy protection.”

^*A separate data source or device that has its own data and its own AI model.