arXiv subjects

Kyung-hwan Lee

Publications and source records attributed to Kyung-hwan Lee.

4 recordsLinked to original sources

arXiv2025

Semantic-Aware Gaussian Process Calibration with Structured Layerwise Kernels for Deep Neural Networks

Calibrating the confidence of neural network classifiers is essential for quantifying the reliability of their predictions during inference. However, conventional Gaussian Process (GP) calibration methods often fail to capture the internal hierarchical structure of deep neural networks, limiting both interpretability and effectiveness for assessing predictive reliability. We propose a Semantic-Aware Layer-wise Gaussian Process (SAL-GP) framework that mirrors the layered architecture of the target neural network. Instead of applying a single global GP correction, SAL-GP employs a multi-layer GP model, where each layer's feature representation is mapped to a local calibration correction. These layerwise GPs are coupled through a structured multi-layer kernel, enabling joint marginalization across all layers. This design allows SAL-GP to capture both local semantic dependencies and global calibration coherence, while consistently propagating predictive uncertainty through the network. The resulting framework enhances interpretability aligned with the network architecture and enables principled evaluation of confidence consistency and uncertainty quantification in deep models.

cs.LG↗

arXiv2024

Adaptive Residual Transformation for Enhanced Feature-Based OOD Detection in SAR Imagery

Recent advances in deep learning architectures have enabled efficient and accurate classification of pre-trained targets in Synthetic Aperture Radar (SAR) images. Nevertheless, the presence of unknown targets in real battlefield scenarios is unavoidable, resulting in misclassification and reducing the accuracy of the classifier. Over the past decades, various feature-based out-of-distribution (OOD) approaches have been developed to address this issue, yet defining the decision boundary between known and unknown targets remains challenging. Additionally, unlike optical images, detecting unknown targets in SAR imagery is further complicated by high speckle noise, the presence of clutter, and the inherent similarities in back-scattered microwave signals. In this work, we propose transforming feature-based OOD detection into a class-localized feature-residual-based approach, demonstrating that this method can improve stability across varying unknown targets' distribution conditions. Transforming feature-based OOD detection into a residual-based framework offers a more robust reference space for distinguishing between in-distribution (ID) and OOD data, particularly within the unique characteristics of SAR imagery. This adaptive residual transformation method standardizes feature-based inputs into distributional representations, enhancing OOD detection in noisy, low-information images. Our approach demonstrates promising performance in real-world SAR scenarios, effectively adapting to the high levels of noise and clutter inherent in these environments. These findings highlight the practical relevance of residual-based OOD detection for SAR applications and suggest a foundation for further advancements in unknown target detection in complex, operational settings.

cs.CV↗

arXiv2023

A search for kilonova radio flares in a sample of Swift/BAT short GRBs

The multi-messenger detection of GW170817 showed that binary neutron star (BNS) mergers are progenitors of (at least some) short gamma-ray bursts (GRBs), and that short GRB jets (and their afterglows) can have structures (and observational properties) more complex than predicted by the standard top-hat jet scenario. Indeed, the emission from the structured jet launched in GW170817 peaked in the radio band (cm wavelengths) at about 100 d since merger - a timescale much longer than the typical time span of radio follow-up observations of short GRBs. Moreover, radio searches for a potential late-time radio flare from the fast tail of the neutron-rich debris that powered the kilonova associated with GW170817 (AT2017gfo) have extended to even longer timescales (years after the merger). In light of this, here we present the results of an observational campaign targeting a sample of seven, years-old GRBs in the Swift/BAT sample with no redshift measurements and no promptly-identified X-ray counterpart. Our goal is to assess whether this sample of short GRBs could harbor nearby BNS mergers, searching for the late-time radio emission expected from their ejecta. We found one radio candidate counterpart for one of the GRBs in our sample, GRB111126A, though an origin related to emission from star formation or from an AGN in its host galaxy cannot be excluded without further observations.

astro-ph.HE↗

arXiv2019

Radio Forensics Could Unmask Nearby Off-axis Gamma-ray Bursts

The multi-messenger observation of gamma-ray burst (GRB)\,170817A from the nearby binary neutron-star merger GW170817 demonstrated that low-energy $γ$-ray emission can be observed at relatively large angles from GRB jet axes. If such structured emission is typical, then the currently known sample of short GRBs with no distance measurements may contain multiple nearby off-axis events whose delayed afterglows could have gone undetected. These nearby neutron star mergers may produce telltale radio flares peaking years after the prompt GRB emission that could still be observable. Here, we show that several short GRBs observed by the Burst Alert Telescope (BAT) on the Neil Gehrels \textit{Swift} satellite, with no identified afterglow and no distance measurement, could potentially be associated with radio flares detectable by sensitive cm-wavelength radio facilities such as the Karl G. Jansky Very Large Array. We also examine optical follow-up observations that have been carried out for these events, and find that a nearby GW170817-like kilonova is ruled out for only a third of them.

astro-ph.HE↗