Selected work

Representative Papers

CTTVAE: Latent Space Structuring for Conditional Tabular Data Generation on Imbalanced Datasets

Feb 03, 2026

This work addresses the challenge of generating high-quality synthetic tabular data for severely imbalanced datasets, where existing methods often fail to preserve both fidelity and utility for minority classes in downstream tasks. The authors propose the CTTVAE+TBS framework, which integrates a conditional Transformer-based variational autoencoder with a class-aware triplet boundary loss to restructure the latent space—enhancing intra-class compactness and inter-class separability. Additionally, an adaptive training sampling mechanism dynamically increases minority class exposure during training. Extensive experiments on six real-world datasets demonstrate that the proposed method significantly outperforms baseline approaches, achieving high data fidelity while substantially improving downstream task performance for minority classes—even surpassing models trained on the original imbalanced data—and effectively bridging the privacy gap between interpolation-based and deep generative methods.

0 citationsRead paper

Multimodal Climate Disinformation Detection: Integrating Vision-Language Models with External Knowledge Sources

Jan 22, 2026

This work addresses the challenge of detecting time-sensitive climate-related multimodal misinformation on social media, which often evades conventional vision-language models due to their reliance on static knowledge acquired during training. To overcome this limitation, the authors propose a novel multimodal detection framework that integrates dynamic external knowledge by incorporating reverse image search, online fact-checking, and authoritative scientific content into the vision-language model. This integration enables real-time retrieval of up-to-date information, thereby enhancing the model’s ability to jointly reason over textual and visual cues in deceptive posts. The proposed approach significantly improves detection accuracy for previously unseen or temporally sensitive climate misinformation, effectively mitigating the constraints imposed by static knowledge bases in existing models.

0 citationsRead paper

Active Inference in Discrete State Spaces from First Principles

Nov 25, 2025

Active inference has long been tightly coupled with the free energy principle (FEP), limiting its conceptual flexibility and interpretability. Method: This work proposes a decoupled framework for active inference in discrete state spaces, reformulating it as a constrained divergence minimization problem—entirely independent of expected free energy. Perception is equivalent to standard variational free energy minimization, while action selection is explicitly optimized via an entropy-regularized objective. The framework integrates mean-field variational inference with information-geometric divergence minimization and embeds probabilistic graphical models to jointly optimize perception and action. Contribution/Results: This is the first formalization of active inference that does not rely on FEP. It preserves behavioral equivalence with classical active inference methods while substantially enhancing interpretability and computational transparency, offering a principled, modular foundation for future theoretical and applied developments in goal-directed behavior modeling.

0 citationsRead paper

SSAVSV: Towards Unified Model for Self-Supervised Audio-Visual Speaker Verification

Jun 21, 2025

Traditional audio-visual speaker verification relies heavily on large-scale labeled data and modality-specific architectures, resulting in high computational overhead and poor generalization. To address these limitations, we propose the first unified self-supervised framework for audio-visual speaker verification: it employs a shared visual Transformer backbone and integrates contrastive learning, asymmetric masking, and masked data modeling to jointly process audio, video, and audio-visual inputs—naturally accommodating missing-modality scenarios. Unlike conventional modality-isolated designs, our approach eliminates the need for labeled data while achieving performance on par with fully supervised methods. It significantly reduces both computational cost and data dependency. Experiments demonstrate strong robustness across multimodal inputs, alongside high efficiency, scalability, and cross-modal consistency. Our framework establishes a novel paradigm for low-resource speaker verification.

0 citationsRead paper

VLM Models and Automated Grading of Atopic Dermatitis

May 23, 2025

Clinical grading of atopic dermatitis (AD) suffers from high subjectivity and low inter-rater agreement among dermatologists. Method: This study systematically evaluates, for the first time, the applicability of seven state-of-the-art vision-language models (VLMs)—including CLIP, Flamingo, and Kosmos-2—to automated, objective AD severity quantification. Leveraging medical domain–specific prompt engineering and zero-shot/few-shot inference, we establish an interpretable, multimodal assessment framework. Results: Multiple VLMs demonstrate superior cross-image stability and clinical plausibility compared to conventional CNNs on public AD datasets. Their severity scores achieve substantial inter-rater agreement with dermatologist panels (Cohen’s κ = 0.72–0.81), validating VLMs as promising tools for objective dermatological quantification. This work provides both methodological innovation and empirical evidence supporting AI-assisted diagnosis and management of skin diseases.

0 citationsRead paper

Recent publications

Latest Papers

CTTVAE: Latent Space Structuring for Conditional Tabular Data Generation on Imbalanced Datasets

Feb 03, 2026

This work addresses the challenge of generating high-quality synthetic tabular data for severely imbalanced datasets, where existing methods often fail to preserve both fidelity and utility for minority classes in downstream tasks. The authors propose the CTTVAE+TBS framework, which integrates a conditional Transformer-based variational autoencoder with a class-aware triplet boundary loss to restructure the latent space—enhancing intra-class compactness and inter-class separability. Additionally, an adaptive training sampling mechanism dynamically increases minority class exposure during training. Extensive experiments on six real-world datasets demonstrate that the proposed method significantly outperforms baseline approaches, achieving high data fidelity while substantially improving downstream task performance for minority classes—even surpassing models trained on the original imbalanced data—and effectively bridging the privacy gap between interpolation-based and deep generative methods.

0 citationsRead paper

Multimodal Climate Disinformation Detection: Integrating Vision-Language Models with External Knowledge Sources

Jan 22, 2026

This work addresses the challenge of detecting time-sensitive climate-related multimodal misinformation on social media, which often evades conventional vision-language models due to their reliance on static knowledge acquired during training. To overcome this limitation, the authors propose a novel multimodal detection framework that integrates dynamic external knowledge by incorporating reverse image search, online fact-checking, and authoritative scientific content into the vision-language model. This integration enables real-time retrieval of up-to-date information, thereby enhancing the model’s ability to jointly reason over textual and visual cues in deceptive posts. The proposed approach significantly improves detection accuracy for previously unseen or temporally sensitive climate misinformation, effectively mitigating the constraints imposed by static knowledge bases in existing models.

0 citationsRead paper

Active Inference in Discrete State Spaces from First Principles

Nov 25, 2025

Active inference has long been tightly coupled with the free energy principle (FEP), limiting its conceptual flexibility and interpretability. Method: This work proposes a decoupled framework for active inference in discrete state spaces, reformulating it as a constrained divergence minimization problem—entirely independent of expected free energy. Perception is equivalent to standard variational free energy minimization, while action selection is explicitly optimized via an entropy-regularized objective. The framework integrates mean-field variational inference with information-geometric divergence minimization and embeds probabilistic graphical models to jointly optimize perception and action. Contribution/Results: This is the first formalization of active inference that does not rely on FEP. It preserves behavioral equivalence with classical active inference methods while substantially enhancing interpretability and computational transparency, offering a principled, modular foundation for future theoretical and applied developments in goal-directed behavior modeling.

0 citationsRead paper

SSAVSV: Towards Unified Model for Self-Supervised Audio-Visual Speaker Verification

Jun 21, 2025

Traditional audio-visual speaker verification relies heavily on large-scale labeled data and modality-specific architectures, resulting in high computational overhead and poor generalization. To address these limitations, we propose the first unified self-supervised framework for audio-visual speaker verification: it employs a shared visual Transformer backbone and integrates contrastive learning, asymmetric masking, and masked data modeling to jointly process audio, video, and audio-visual inputs—naturally accommodating missing-modality scenarios. Unlike conventional modality-isolated designs, our approach eliminates the need for labeled data while achieving performance on par with fully supervised methods. It significantly reduces both computational cost and data dependency. Experiments demonstrate strong robustness across multimodal inputs, alongside high efficiency, scalability, and cross-modal consistency. Our framework establishes a novel paradigm for low-resource speaker verification.

0 citationsRead paper

VLM Models and Automated Grading of Atopic Dermatitis

May 23, 2025

Clinical grading of atopic dermatitis (AD) suffers from high subjectivity and low inter-rater agreement among dermatologists. Method: This study systematically evaluates, for the first time, the applicability of seven state-of-the-art vision-language models (VLMs)—including CLIP, Flamingo, and Kosmos-2—to automated, objective AD severity quantification. Leveraging medical domain–specific prompt engineering and zero-shot/few-shot inference, we establish an interpretable, multimodal assessment framework. Results: Multiple VLMs demonstrate superior cross-image stability and clinical plausibility compared to conventional CNNs on public AD datasets. Their severity scores achieve substantial inter-rater agreement with dermatologist panels (Cohen’s κ = 0.72–0.81), validating VLMs as promising tools for objective dermatological quantification. This work provides both methodological innovation and empirical evidence supporting AI-assisted diagnosis and management of skin diseases.

0 citationsRead paper