Selected work

Representative Papers

DAVE: Distribution-aware Attribution via ViT Gradient Decomposition

Feb 06, 2026

This work addresses the limitations of attribution maps generated by Vision Transformers (ViTs), which are often compromised by structured artifacts stemming from patch embeddings and attention mechanisms, yielding only coarse-grained and unstable block-level explanations. To overcome this, the authors propose a gradient decomposition method tailored to the architectural characteristics of ViTs, introducing— for the first time—distribution-aware modeling to mathematically disentangle the local, equivariant, and stable input–output mapping components from structural noise. This approach effectively suppresses architecture-induced artifacts, substantially enhancing both the stability and spatial resolution of attributions, and thereby producing high-fidelity, pixel-level explanation maps.

0 citationsRead paper

ProtoQuant: Quantization of Prototypical Parts For General and Fine-Grained Image Classification

Feb 06, 2026

This work addresses the limitations of prototype-based models in large-scale image classification—namely, their weak generalization, reliance on costly fine-tuning, and susceptibility to prototype drift—by introducing vector quantization into the latent space. The proposed approach employs a discrete, learnable codebook to constrain prototype representations, enabling stable, data-anchored, and interpretable prototype modeling without requiring fine-tuning of the backbone network. Evaluated on benchmark datasets including ImageNet, CUB-200, and Cars-196, the method achieves competitive classification accuracy while significantly enhancing model interpretability and prototype consistency.

0 citationsRead paper

SIDE: Sparse Information Disentanglement for Explainable Artificial Intelligence

Jul 25, 2025

To address the redundancy and poor interpretability of prototypical neural networks in large-scale vision tasks (e.g., ImageNet), this paper proposes a sparse information disentanglement mechanism. Within the prototypical network framework, we introduce sparsity-inducing regularization during training, replace the softmax classifier with a sigmoid-based class-level activation function, and design a prototype pruning strategy that enforces each class decision to rely only on a small subset of salient prototypes. The method significantly enhances conceptual clarity and interpretability at the prototype level: it compresses the explanation size by over 90% while preserving model accuracy—substantially outperforming baselines such as InfoDisent. Our core contribution is the first effective, sparse, and semantically transparent prototype-level interpretability model tailored for large-scale datasets.

0 citationsRead paper

Personalized Interpretability -- Interactive Alignment of Prototypical Parts Networks

Jun 05, 2025

Concept-based explainable neural networks suffer from concept inconsistency—e.g., conflating bird heads and wings into a single concept—leading to explanations misaligned with human cognition; moreover, existing methods lack mechanisms to incorporate users’ personalized preferences regarding concept appearance. This paper introduces the first user-driven concept adaptive segmentation framework, embedding interactive feedback directly into the prototype learning process of ProtoPNet. Through user-annotated guidance, the method enables concept splitting, re-clustering, and consistency regularization, jointly optimizing both interpretability and individual cognitive alignment. Evaluated on FunnyBirds, CUB, CARS, and PETS, our approach achieves significant improvements in concept consistency, boosts user satisfaction by 42%, and maintains classification accuracy without degradation.

0 citationsRead paper

B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data

May 28, 2025

Existing GNN interpretability evaluation frameworks heavily rely on synthetic data and proxy metrics, failing to reflect the faithfulness of explanations in real-world chemical scenarios. Method: We propose B-XAIC—the first XAI benchmark for molecular graphs—built on real molecular datasets and tasks, incorporating expert-validated ground-truth rationale annotations. Our methodology integrates molecular graph representation learning, systematic evaluation of GNN explanation algorithms, domain-expert annotation, and a rigorous consistency verification protocol. Contribution/Results: Experiments reveal widespread faithfulness deficits among mainstream GNN explanation methods on chemical tasks. B-XAIC establishes the first rationale-driven, chemistry-specific interpretability evaluation paradigm, providing a standardized, reproducible testing platform for XAI algorithms. It significantly enhances model trustworthiness and debuggability in molecular machine learning, enabling rigorous, domain-grounded assessment of explanation quality.

0 citationsRead paper

Recent publications

Latest Papers

DAVE: Distribution-aware Attribution via ViT Gradient Decomposition

Feb 06, 2026

This work addresses the limitations of attribution maps generated by Vision Transformers (ViTs), which are often compromised by structured artifacts stemming from patch embeddings and attention mechanisms, yielding only coarse-grained and unstable block-level explanations. To overcome this, the authors propose a gradient decomposition method tailored to the architectural characteristics of ViTs, introducing— for the first time—distribution-aware modeling to mathematically disentangle the local, equivariant, and stable input–output mapping components from structural noise. This approach effectively suppresses architecture-induced artifacts, substantially enhancing both the stability and spatial resolution of attributions, and thereby producing high-fidelity, pixel-level explanation maps.

0 citationsRead paper

ProtoQuant: Quantization of Prototypical Parts For General and Fine-Grained Image Classification

Feb 06, 2026

This work addresses the limitations of prototype-based models in large-scale image classification—namely, their weak generalization, reliance on costly fine-tuning, and susceptibility to prototype drift—by introducing vector quantization into the latent space. The proposed approach employs a discrete, learnable codebook to constrain prototype representations, enabling stable, data-anchored, and interpretable prototype modeling without requiring fine-tuning of the backbone network. Evaluated on benchmark datasets including ImageNet, CUB-200, and Cars-196, the method achieves competitive classification accuracy while significantly enhancing model interpretability and prototype consistency.

0 citationsRead paper

SIDE: Sparse Information Disentanglement for Explainable Artificial Intelligence

Jul 25, 2025

To address the redundancy and poor interpretability of prototypical neural networks in large-scale vision tasks (e.g., ImageNet), this paper proposes a sparse information disentanglement mechanism. Within the prototypical network framework, we introduce sparsity-inducing regularization during training, replace the softmax classifier with a sigmoid-based class-level activation function, and design a prototype pruning strategy that enforces each class decision to rely only on a small subset of salient prototypes. The method significantly enhances conceptual clarity and interpretability at the prototype level: it compresses the explanation size by over 90% while preserving model accuracy—substantially outperforming baselines such as InfoDisent. Our core contribution is the first effective, sparse, and semantically transparent prototype-level interpretability model tailored for large-scale datasets.

0 citationsRead paper

Personalized Interpretability -- Interactive Alignment of Prototypical Parts Networks

Jun 05, 2025

Concept-based explainable neural networks suffer from concept inconsistency—e.g., conflating bird heads and wings into a single concept—leading to explanations misaligned with human cognition; moreover, existing methods lack mechanisms to incorporate users’ personalized preferences regarding concept appearance. This paper introduces the first user-driven concept adaptive segmentation framework, embedding interactive feedback directly into the prototype learning process of ProtoPNet. Through user-annotated guidance, the method enables concept splitting, re-clustering, and consistency regularization, jointly optimizing both interpretability and individual cognitive alignment. Evaluated on FunnyBirds, CUB, CARS, and PETS, our approach achieves significant improvements in concept consistency, boosts user satisfaction by 42%, and maintains classification accuracy without degradation.

0 citationsRead paper

B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data

May 28, 2025

Existing GNN interpretability evaluation frameworks heavily rely on synthetic data and proxy metrics, failing to reflect the faithfulness of explanations in real-world chemical scenarios. Method: We propose B-XAIC—the first XAI benchmark for molecular graphs—built on real molecular datasets and tasks, incorporating expert-validated ground-truth rationale annotations. Our methodology integrates molecular graph representation learning, systematic evaluation of GNN explanation algorithms, domain-expert annotation, and a rigorous consistency verification protocol. Contribution/Results: Experiments reveal widespread faithfulness deficits among mainstream GNN explanation methods on chemical tasks. B-XAIC establishes the first rationale-driven, chemistry-specific interpretability evaluation paradigm, providing a standardized, reproducible testing platform for XAI algorithms. It significantly enhances model trustworthiness and debuggability in molecular machine learning, enabling rigorous, domain-grounded assessment of explanation quality.

0 citationsRead paper