Institution profile

Ulsan National Institute of Science and Technology

Academic institutionasia · kr
Official website
Research library239linked papers
Opportunities0open roles
Selected work

Representative Papers

Conditional Temporal Neural Processes with Covariance Loss

Apr 01, 2025International Conference on Machine Learning

Neural processes often fail to accurately model dependencies between inputs and targets under noisy observations. To address this, we propose the Covariance Loss—a novel objective that explicitly incorporates second-order statistical dependencies among target variables into the end-to-end training of conditional neural processes for the first time. By regularizing the covariance structure of the predictive distribution, our loss enhances the model’s ability to recover missing or degraded dependencies and improves robustness to observation noise. The method is architecture-agnostic and can be seamlessly integrated into mainstream neural process frameworks. Extensive experiments across multiple real-world time-series and regression benchmarks demonstrate consistent and significant improvements over state-of-the-art methods in three key aspects: predictive accuracy, fidelity of dependency structure recovery, and robustness to observational noise.

15 citationsRead paper

Stable Neural Stochastic Differential Equations in Analyzing Irregular Time Series Data

Feb 22, 2024International Conference on Learning Representations

Real-world time series (e.g., stock prices, meteorological data) often exhibit irregular sampling and pervasive missing values, posing fundamental challenges for continuous-time modeling and robust inference. Method: We propose three theoretically guaranteed stable neural stochastic differential equations (Neural SDEs): Langevin-type, linear-noise-type, and geometric-type. Grounded in Itô calculus, our framework rigorously ensures strong solution existence, numerical stability, and distributional shift robustness. We further design an adaptive Euler–Maruyama solver and a missingness-aware likelihood optimization scheme to enable continuous latent-space modeling and robust inference. Results: Our approach achieves significant improvements over baselines—including Neural ODEs and GRU-D—across interpolation, forecasting, and classification tasks on four benchmark datasets. Moreover, extensive evaluation across 30 public datasets under diverse missingness rates confirms consistent generalization performance and strong robustness to missing data.

8 citationsRead paper

Flow Actor-Critic for Offline Reinforcement Learning

Feb 20, 2026

This work addresses the challenge in offline reinforcement learning that multimodal data distributions cannot be effectively modeled by conventional Gaussian policies. To this end, the authors propose Flow Actor-Critic (FAC), a novel method that, for the first time, integrates normalizing flows into both the policy network and the conservative critic design. Specifically, the approach leverages flow-based models to construct a highly expressive policy capable of accurately capturing complex behavioral distributions. Furthermore, it introduces a flow-based behavioral proxy model to derive a new regularizer for the critic, thereby enhancing the conservatism of value estimation. The proposed method achieves state-of-the-art performance on established offline reinforcement learning benchmarks, including D4RL and OGBench.

1 citationsRead paper

DualDynamics: Synergizing Implicit and Explicit Methods for Robust Irregular Time Series Analysis

Jan 10, 2024

To address the challenges of time-series modeling under irregular sampling, high missingness rates, and strong temporal heterogeneity, this paper proposes the first dual-dynamics framework that jointly integrates neural ordinary differential equations (implicit dynamics) with neural flows (explicit dynamics). Our approach employs adaptive interpolation-based encoding and a joint implicit–explicit gradient optimization scheme, ensuring numerical stability while significantly enhancing generalization to distribution shifts and partial observations. The method achieves state-of-the-art performance across multiple tasks—including classification, imputation, and forecasting—reducing average error by 23.6% under high missingness and strong temporal heterogeneity. Crucially, it harmonizes expressive power, numerical stability, and computational efficiency, offering a principled solution for robust time-series representation learning in challenging real-world settings.

1 citationsRead paper
Recent publications

Latest Papers