This inventor holds 1 USPTO granted patent. Top assignee: Asapp, Inc.. Active years: 2022.
Data source: USPTO (United States Patent and Trademark Office) public patent records. Weekly synchronization. How IDiyas builds this profile
Company Filing History:
Years Active: 2022
Title: Keunwoo Peter Yu: Advancing Named Entity Recognition Through Marginal Distillation.
Introduction
Keunwoo Peter Yu is an inventor based in Detroit, Michigan. His patent work addresses how artificial intelligence models can combine knowledge from multiple named entity recognition systems.
Latest Patents
Yu is associated with 1 patent: “System, Method, and Computer Program for Obtaining a Unified Named Entity Recognition Model with the Collective Predictive Capabilities of Teacher Models with Different Tag Sets Using Marginal Distillation.”
The invention describes a method for training a unified student model from multiple pre-trained teacher models that use different tag sets. It distills knowledge without requiring access to the annotated training data used to train the teachers. A tag hierarchy brings the different tag sets together, and a distillation loss compares the student’s predictions with those of each teacher. When a teacher’s predictions are less fine-grained than the student’s, the method marginalizes the student’s more detailed predictions for the relevant hierarchy node. Minimizing the overall loss enables the student model to acquire the teachers’ combined predictive capabilities.
Career Highlights
Yu is associated with Asapp, Inc. His patent reflects work on machine learning methods for consolidating information from models with differing classification schemes.
Collaborations
Yu’s listed coworker is Yi Yang. The patent’s approach also involves the collective knowledge of multiple teacher models, which contribute their predictive capabilities to a unified student model.
Conclusion
Keunwoo Peter Yu’s patent presents a way to combine named entity recognition models with different tag sets into a unified system, while preserving their collective predictive strengths.
Data source: USPTO (United States Patent and Trademark Office) public patent records. Weekly synchronization. How IDiyas builds this profile
