Toss Paper Accepted at World’s Top AI Conference 'NeurIPS 2025'

Toss Achieves Global Recognition for Federated Learning Optimization
Research Led by Jinwoo Lee Accepted at NeurIPS 2025
New Algorithm Enables Privacy-Compliant AI Training Across Diverse Data Environments

Viva Republica, the operator of Toss, announced on December 3 that its research paper has been accepted for presentation at NeurIPS 2025, the world's most prestigious artificial intelligence (AI) conference.

The paper is the result of a collaborative research project led by Jinwoo Lee, a researcher on the Toss Face Modeling Team, in partnership with the Vision Lab at Seoul National University.

According to Toss, NeurIPS is the most influential conference in the field of machine learning and neural information processing worldwide, with a paper acceptance rate of about 20 percent. The conference is being held at the San Diego Convention Center in the United States from the previous day through December 7.

The accepted research, titled "Federated Learning with Local Prior Alignment (FedLPA)," enables AI models to be trained even in countries where data cannot be transferred to a central server due to privacy protection regulations.

This approach addresses the limitations of existing federated learning methods, which experience significant performance drops when data characteristics differ across countries or user groups, or when previously unseen types of data emerge.

The research team combined local clustering based on "Infomap," which automatically groups data with similar characteristics such as country or user group, with a "local prior alignment" technique that aligns predictions to enhance learning stability.

The method allows each device to autonomously identify and utilize its own data structure, thereby demonstrating "generalized category discovery (GCD)" performance by accurately detecting new types of categories even in environments where it is difficult to know the number of categories or the data distribution in advance.

Through this research, the team demonstrated that it is possible to build global AI models that comply with the legal requirements of countries with strict privacy regulations, fulfilling both privacy protection and AI performance as a foundational technology.

The title of the paper is "Local Prior Alignment for Generalized Category Discovery in Heterogeneous Federated Learning Environments (FedLPA)."

Researcher Lee stated, "The core of this research is the optimization of the algorithm to enable efficient learning even in challenging situations where data cannot be transferred to a server due to regulations, client data distributions are highly diverse, and even the number of new categories is unknown."

A Toss representative commented, "It is highly significant that Toss's AI capabilities have received official recognition from a global academic conference for the first time," adding, "We will continue to pursue research into technologies that can be applied to our services, providing more sophisticated AI-based services while safeguarding privacy protection."

This content was produced with the assistance of AI translation services.