该研究通过引入Modality Contribution Network (MCNet) 和 Modality Contribution Score (MCS)，量化了在阿尔茨海默病诊断中结构MRI和淀粉样PET的相对贡献，解决了现有AI系统无法为每个患者提供具体模态贡献的问题。
This study addresses the $q$-Bin Packing ($q$BP) problem with item replication and placement constraints, where the objective is to minimize the number of bins used such that each bin contains at most one copy of any item and respects a given capacity limit. Focusing on the generalized First Fit Decreasing (FFD) algorithm, the paper establishes, for the first time, a performance guarantee for $\mathsf{FFD}_q$ in a special case of $q$BP without relying on the "single-item-at-the-end" assumption. By constructing a carefully designed subinstance ${D'}_q$ and leveraging techniques from combinatorial optimization and approximation analysis, the authors prove that $\mathsf{FFD}_q(D_q) \leq \frac{11}{9} \mathsf{OPT}(D_q) + 3q$. This result provides the first effective approximation ratio bound for this variant of the bin packing problem.
This work addresses the challenge of evaluating whether deep image classification models rely on task-relevant regions, given their opaque decision-making processes. To this end, the authors propose ReMoDEx, a novel framework that integrates local attribution methods—such as GradCAM++ and Integrated Gradients—with global heatmap clustering to enable systematic analysis from sample-level explanations to dataset-level decision patterns. By standardizing heatmaps, performing similarity-based clustering, and assessing spatial correlations, ReMoDEx uncovers shortcut learning behaviors invisible to conventional evaluation metrics. Applied to a COVID-19 chest X-ray classification task, the framework reveals two dominant model strategies: reliance on either the central thoracic region or image borders. Occlusion experiments confirm the latter as a shortcut, despite the model achieving a test accuracy of 86.27% and an AUC of 0.9624.
This study addresses the limited generalization of existing voice deepfake detection methods on Southeast Asian (SEA) languages and the absence of a multilingual benchmark for neural audio codec–generated synthetic speech (Codecfakes). To bridge this gap, we introduce SEA-CF, the first large-scale Codecfakes detection benchmark tailored to SEA languages, encompassing diverse languages, speakers, and neural audio codecs. We also propose GARUDA, a lightweight audio language model specifically designed for low-resource and low-latency deployment scenarios. Experimental results demonstrate that state-of-the-art English-centric detectors suffer significant performance degradation on SEA languages, whereas GARUDA outperforms both end-to-end and large audio language model baselines on SEA-CF, achieving high accuracy while maintaining practical deployability.
This study presents the first systematic evaluation of instance-level machine unlearning methods for 3D medical image segmentation, addressing the need to efficiently remove data from specific individuals in compliance with privacy regulations such as GDPR. Leveraging a Med3D-pretrained 3D ResNet-50 backbone, four approximate unlearning strategies are compared on the MRBrainS18 dataset, with performance trade-offs between forgetting efficacy and retained utility quantified using Dice coefficient and mean absolute error (MAE). Experimental results demonstrate that the Noisy Label strategy reduces performance on the forget set by 93% after 50 training epochs while maintaining 84% accuracy on the retain set—significantly outperforming other approaches, which all suffer catastrophic degradation in retained performance under prolonged training. This work establishes a rigorous baseline for the task and identifies an optimal balance between effective unlearning and model utility.
该研究通过引入Modality Contribution Network (MCNet) 和 Modality Contribution Score (MCS)，量化了在阿尔茨海默病诊断中结构MRI和淀粉样PET的相对贡献，解决了现有AI系统无法为每个患者提供具体模态贡献的问题。
This study addresses the $q$-Bin Packing ($q$BP) problem with item replication and placement constraints, where the objective is to minimize the number of bins used such that each bin contains at most one copy of any item and respects a given capacity limit. Focusing on the generalized First Fit Decreasing (FFD) algorithm, the paper establishes, for the first time, a performance guarantee for $\mathsf{FFD}_q$ in a special case of $q$BP without relying on the "single-item-at-the-end" assumption. By constructing a carefully designed subinstance ${D'}_q$ and leveraging techniques from combinatorial optimization and approximation analysis, the authors prove that $\mathsf{FFD}_q(D_q) \leq \frac{11}{9} \mathsf{OPT}(D_q) + 3q$. This result provides the first effective approximation ratio bound for this variant of the bin packing problem.
This work addresses the challenge of evaluating whether deep image classification models rely on task-relevant regions, given their opaque decision-making processes. To this end, the authors propose ReMoDEx, a novel framework that integrates local attribution methods—such as GradCAM++ and Integrated Gradients—with global heatmap clustering to enable systematic analysis from sample-level explanations to dataset-level decision patterns. By standardizing heatmaps, performing similarity-based clustering, and assessing spatial correlations, ReMoDEx uncovers shortcut learning behaviors invisible to conventional evaluation metrics. Applied to a COVID-19 chest X-ray classification task, the framework reveals two dominant model strategies: reliance on either the central thoracic region or image borders. Occlusion experiments confirm the latter as a shortcut, despite the model achieving a test accuracy of 86.27% and an AUC of 0.9624.
This study addresses the limited generalization of existing voice deepfake detection methods on Southeast Asian (SEA) languages and the absence of a multilingual benchmark for neural audio codec–generated synthetic speech (Codecfakes). To bridge this gap, we introduce SEA-CF, the first large-scale Codecfakes detection benchmark tailored to SEA languages, encompassing diverse languages, speakers, and neural audio codecs. We also propose GARUDA, a lightweight audio language model specifically designed for low-resource and low-latency deployment scenarios. Experimental results demonstrate that state-of-the-art English-centric detectors suffer significant performance degradation on SEA languages, whereas GARUDA outperforms both end-to-end and large audio language model baselines on SEA-CF, achieving high accuracy while maintaining practical deployability.
This study presents the first systematic evaluation of instance-level machine unlearning methods for 3D medical image segmentation, addressing the need to efficiently remove data from specific individuals in compliance with privacy regulations such as GDPR. Leveraging a Med3D-pretrained 3D ResNet-50 backbone, four approximate unlearning strategies are compared on the MRBrainS18 dataset, with performance trade-offs between forgetting efficacy and retained utility quantified using Dice coefficient and mean absolute error (MAE). Experimental results demonstrate that the Noisy Label strategy reduces performance on the forget set by 93% after 50 training epochs while maintaining 84% accuracy on the retain set—significantly outperforming other approaches, which all suffer catastrophic degradation in retained performance under prolonged training. This work establishes a rigorous baseline for the task and identifies an optimal balance between effective unlearning and model utility.