Saerin Lim
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Saerin Lim has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Seoung Bum Kim
Korea University
5 shared publications Weon Gyu Shin
3 shared publications- Jinsoo Bae
Korea University
2 shared publications - Jongkook Heo
Korea University
2 shared publications Yu Zhang
2 shared publicationsVan Tuan Ho
1 shared publication
Latest Publications
- Open Access
- 3 August 2026
- Published by Elsevier in Nuclear Engineering and Technology
- Vol. 59 (1)
An uncertainty-aware machine learning framework for predicting cable and cabinet damage in switchgear room fire scenarios
- 3 August 2026
- Published by Elsevier in Nuclear Engineering and Technology
- Vol. 59 (1)
Open AccessAn uncertainty-aware machine learning framework for predicting cable and cabinet damage in switchgear room fire scenarios
- Open Access
- 8 July 2026
- Published by Elsevier in SSRN Electronic Journal
OwareMatch: Transferring Open-Set Knowledge for Open-Set Semi-Supervised Learning
Open-set semi-supervised learning (OSSL) addresses the practical scenario in which unlabeled data contains out-of-distribution (OOD) samples. Early OSSL methods typ- ically relied on strict hard-filtering, which frequently led to severe over-rejection of valid samples and poor precision. To mitigate this, recent approaches have adopted soft-filtering or explicit outlier modeling; however, these methods often merely shift the failure mode along a pronounced precision-recall trade-off. Specifically, methods that learn a dedicated unknown-class representation overfit to seen OOD patterns, while soft-filtering methods suffer from an ID-dominated optimization where outlier scores only attenuate gradient magnitude. Because this magnitude-only reduction fails to alter the optimization direction, OOD updates are progressively absorbed into known-class regions, ultimately leading to poor recall and under-rejection. To overcome this opti- mization limitation, we propose OwareMatch, which integrates open-set knowledge directly into the classifier’s gradient flow. By decoupling logits into ID and OOD energy components and asymmetrically modulating gradients, OwareMatch progres- sively decorrelates the ID and OOD optimization paths. This gradient-level decoupling suppresses OOD absorption into ID decision boundaries, instead encouraging OOD features toward a class-agnostic rejection region. Because this rejection mechanism avoids explicitly learned, seen OOD-specific representations, it is less distribution- dependent and transfers more robustly to unseen OOD shifts. Extensive experiments demonstrate that OwareMatch effectively mitigates the precision-recall trade-off, main- taining competitive closed-set accuracy while consistently improving OOD F1-score and robustness under both seen and unseen shifts.
- 8 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessOwareMatch: Transferring Open-Set Knowledge for Open-Set Semi-Supervised Learning
Open-set semi-supervised learning (OSSL) addresses the practical scenario in which unlabeled data contains out-of-distribution (OOD) samples. Early OSSL methods typ- ically relied on strict hard-filtering, which frequently led to severe over-rejection of valid samples and poor precision. To mitigate this, recent approaches have adopted soft-filtering or explicit outlier modeling; however, these methods often merely shift the failure mode along a pronounced precision-recall trade-off. Specifically, methods that learn a dedicated unknown-class representation overfit to seen OOD patterns, while soft-filtering methods suffer from an ID-dominated optimization where outlier scores only attenuate gradient magnitude. Because this magnitude-only reduction fails to alter the optimization direction, OOD updates are progressively absorbed into known-class regions, ultimately leading to poor recall and under-rejection. To overcome this opti- mization limitation, we propose OwareMatch, which integrates open-set knowledge directly into the classifier’s gradient flow. By decoupling logits into ID and OOD energy components and asymmetrically modulating gradients, OwareMatch progres- sively decorrelates the ID and OOD optimization paths. This gradient-level decoupling suppresses OOD absorption into ID decision boundaries, instead encouraging OOD features toward a class-agnostic rejection region. Because this rejection mechanism avoids explicitly learned, seen OOD-specific representations, it is less distribution- dependent and transfers more robustly to unseen OOD shifts. Extensive experiments demonstrate that OwareMatch effectively mitigates the precision-recall trade-off, main- taining competitive closed-set accuracy while consistently improving OOD F1-score and robustness under both seen and unseen shifts.
- Open Access
- 29 October 2025
- Published by Elsevier in Nuclear Engineering and Technology
- Vol. 58 (3)
CFAST simulations and application of a machine learning algorithm for the fire safety of a switchgear room
- 29 October 2025
- Published by Elsevier in Nuclear Engineering and Technology
- Vol. 58 (3)
Open AccessCFAST simulations and application of a machine learning algorithm for the fire safety of a switchgear room
- Open Access
- 19 August 2025
- Published by Elsevier in Nuclear Engineering and Technology
- Vol. 58 (1)
Explainable machine learning-based meta-modeling for predicting fire damage to motor control cabinets in a switchgear room
- 19 August 2025
- Published by Elsevier in Nuclear Engineering and Technology
- Vol. 58 (1)
Open AccessExplainable machine learning-based meta-modeling for predicting fire damage to motor control cabinets in a switchgear room
- 2 February 2024
- Published by Elsevier in Knowledge-Based Systems
- Vol. 287
SWISS: Signal-wise self-supervised learning to learn a multi-signal interaction for human activity recognition
- 2 February 2024
- Published by Elsevier in Knowledge-Based Systems
- Vol. 287
SWISS: Signal-wise self-supervised learning to learn a multi-signal interaction for human activity recognition