Publications
Peer-reviewed journal and conference papers.
2025
-
Multimodal Alzheimer's disease recognition from image, text and audio
Builds an image-sentence bipartite graph with a vision-language model and fuses the resulting GCN representation with text and speech encoders through co-attention, then analyses each modality's contribution.
Key result. On the ADReSSo benchmark, fusing the graph with text and audio through co-attention improved mean accuracy over the graph alone; most of the gain came from co-attention across modalities, and adding audio was not beneficial in every configuration. Results are for an English picture-description benchmark, not a clinical validation.
-
Alzheimer's disease recognition using graph neural network by leveraging image-text similarity from vision language model
Represents picture regions and spoken sentences in a shared vision-language space and classifies participants from the similarity-weighted bipartite graph, with ablations on the relational structure.
Key result. On the ADReSSo benchmark, the graph built from image-text similarity supported Alzheimer's disease classification on its own, and removing the image-text relations lowered mean accuracy, indicating that the relational structure itself carries diagnostic signal.
2023
-
Burst and Memory-aware Transformer: capturing temporal heterogeneity
Embeds burstiness and memory statistics into a Transformer temporal point process and tests where explicit temporal heterogeneity helps next-event time prediction.
Key result. Burst and memory statistics improved next-event time prediction mainly where temporal heterogeneity was strong; against a matched ablation the gain held on all but one of the real-world datasets.
-
Latent-Variable Classifiers Based on Total Correlation for Dynamic Environments
Couples total-correlation latent variables with existing classifiers so that models can absorb newly arriving features without revisiting past raw data.
-
Explainable ML models for a deeper insight on treatment decision for localized prostate cancer
Combines registry and regional health-resource data to model observed treatment choices for localized prostate cancer, using SHAP to examine how clinical and socioeconomic factors interact.
2022
-
Brain-inspired Predictive Coding Improves the Performance of Machine Challenging Tasks
Applies predictive-coding learning to incremental, long-tailed and few-shot recognition, with comparisons against backpropagation-based learning.
2019
-
Copula-based algorithm for generating bursty time series
Uses a Farlie-Gumbel-Morgenstern copula to generate event sequences with a prescribed inter-event time distribution and a prescribed memory coefficient within the correlation range allowed by the construction.
2018
-
Hierarchical burst model for complex bursty dynamics
Proposes a hierarchical generative model in which one mechanism accounts for the inter-event time distribution, the autocorrelation and the burst size distribution at once.
Key result. The relation between the inter-event time exponent and the autocorrelation exponent (alpha + gamma = 2) still holds when successive intervals are correlated, and burst sizes follow a stretched-exponential rather than a power-law distribution.
-
Analysis on the urban street network of Korea: Connections between topology and meta-information
Analyses the street networks of 22 Korean cities through centrality measures, planning regularity and population-road length scaling, and relates the structural groups to urban meta-information.
Key result. Cities grouped by the inequality of their centrality measures shared demographic and planning characteristics, and population-road-length scaling differed between urban and urban-rural cities.
Demonstrations and other contributions
-
Intelligent System for Early Detection of Mild Cognitive Impairment and Dementia Risk Through Spoken Language Analysis
A screening system for cognitive-decline risk based on spoken language analysis.
-
DementiaBank Korean Kang Corpus