UNIST Research Team Develops Training Technology to Prevent Performance Degradation in AI Models

Robust Time Series Learning Techniques Against Data Drift

"Theoretical and Experimental Validation Completed... Increased Potential for AI Applications"

A learning technology that can effectively respond to the phenomenon that degrades the performance of artificial intelligence models has been developed. This research is expected to contribute to enhancing the applicability and performance of AI in industrial fields.

UNIST (President Yong-Hoon Lee) announced on the 25th that a research team led by Professors Sungil Kim and Dongyoung Lim from the Department of Industrial Engineering and the Graduate School of Artificial Intelligence developed a "time series learning technology robust to data drift."

Time series data refers to data collected continuously with a certain periodicity in chronological order. Numerous data used in various industries such as finance, economy, transportation, agriculture, manufacturing, and healthcare are in time series form.

Time series data experiences a phenomenon called "data drift" as external factors influencing data generation change. Data drift refers to the difference between the data used for training AI models and the actual data in the operational environment.

(From left) Professor Kim Seong-il, Professor Im Dong-young, First author Researcher Oh Yong-kyung.

Professor Sungil Kim stated, "When data drift occurs, the performance of time series learning AI models deteriorates," adding, "It is a chronic problem that makes it difficult to utilize time series data in various industries."

The research team developed a methodology for designing a robust neural network structure based on Neural SDEs (Stochastic Differential Equations) that can effectively respond to this problem.

Neural SDEs are an extended model of Neural ODEs, which are continuous versions of residual neural network models. The team presented theoretical grounds for designing time series Neural SDEs models that maintain robustness even under data drift phenomena.

Following the methodology, the team introduced three Neural SDEs models: Langevin-type SDE, Linear Noise SDE, and Geometric SDE. The proposed models demonstrated stable and superior performance in various tasks such as interpolation, prediction, and classification on datasets experiencing data drift.

The engineering process of quickly detecting data drift, reconstructing data, and retraining involved significant time and cost.

The research team has theoretically and experimentally verified the technology necessary to make AI robust to data drift from the outset.

Professor Dongyoung Lim explained, "Recently, cases of performance degradation in time series AI models due to data drift in dynamic data environments have become frequent," adding, "This research is significant in that it developed a methodology to train AI to be robust to drift from the beginning and verified its performance both theoretically and experimentally."

First author Researcher Yongkyung Oh said, "We developed a neural network structure design methodology to prevent AI performance degradation caused by time series data drift," and added, "We plan to continuously develop time series data drift monitoring technology and learning data reconstruction technology to enable various domestic companies to utilize them."

This research was selected as a spotlight paper, ranking in the top 5% at the internationally prestigious conference ICLR (International Conference on Learning Representations), and will be presented in May in Vienna, Austria.

It was supported by the Korea Health Industry Development Institute's Biomedical Global Talent Training Project, the Institute for Information & Communications Technology Planning & Evaluation's AI Graduate School Project, and the Ministry of Science and ICT's National Research Foundation Basic Research and Human-Centered Carbon Neutral Global Supply Chain Research Center.

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