Selected work

Representative Papers

CalTwin: Towards Calibrated, Shift-Robust Medical World Models via Fisher-Information Regularisation

Jul 29, 2026

This work addresses covariate shift and confidence miscalibration in clinical deployment of medical world models by proposing CalTwin, a lightweight regularization method that uniquely integrates Fisher Information Matrix (FIM) penalization with confidence miscalibration penalization into a GRU-based latent state transition predictor. Through joint optimization, CalTwin simultaneously mitigates performance degradation under out-of-distribution data and overconfidence. Evaluated on the PhysioNet 2019 Sepsis Challenge dataset, CalTwin reduces the mean squared error of out-of-distribution next-step latent state prediction by 9.1%—with 7.0% attributable to FIM regularization—and modestly improves calibration, lowering expected calibration error (ECE) by 0.7%, thereby significantly enhancing model reliability and calibration in multi-center, heterogeneous clinical settings.

0 citationsRead paper

PIcsC: Partitioning-Induced Covariate Shift Correction

Jul 28, 2026

This work addresses covariate shift induced by data partitioning, which severely undermines model selection and evaluation in cross-validation, continual learning, and federated learning. The authors propose the PIcsC framework, which introduces a Fisher information–based regularization term to uniformly mitigate partition-induced shifts in both centralized and distributed settings. PIcsC approximates distributional discrepancies—detected via KL divergence—using local gradient statistics and activates conditional adaptive regularization only when significant shifts are identified, without requiring access to raw data. Evaluated across more than 40 datasets, PIcsC substantially alleviates performance degradation, reducing loss by 20% under batch partitioning and by 25% in cross-validation. Moreover, it outperforms baselines such as FedAvg by 3–5 percentage points on seven federated learning benchmarks.

0 citationsRead paper

Technical note on Sequential Test-Time Adaptation via Martingale-Driven Fisher Prompting

Oct 04, 2025

This paper addresses robust sequential decision-making under covariate shift in streaming data. Methodologically, it proposes the M-FISHER framework, which unifies real-time distributional shift detection and stable online adaptation. It constructs an exponential martingale from non-conformity scores and leverages Ville’s inequality to guarantee bounded false alarm rate at any time; additionally, it employs Fisher-preconditioned parameter updates to perform natural gradient descent on the distribution manifold, ensuring geometric invariance and KL-divergence minimization. Theoretically, it is the first to integrate martingale theory with Fisher information geometry for test-time adaptation, delivering time-uniform statistical guarantees; it also derives an explicit upper bound on detection delay—O(log(1/δ)/Γ)—that scales inversely with the magnitude Γ of distributional discrepancy. Experiments demonstrate its superior detection efficiency and adaptive stability.

0 citationsRead paper

Technical note on Fisher Information for Robust Federated Cross-Validation

Oct 04, 2025

In federated learning, spatiotemporally distributed data induce covariate shift, causing local empirical distributions across clients to deviate from the global underlying distribution and thereby degrading model generalization. To address this, we propose FIRE (Fisher Information Regularized Estimation), the first method to incorporate the Fisher information matrix into federated cross-validation. FIRE approximates the Fisher information distance between each client’s local data and the global distribution, quantifying and correcting covariate shift. This distance is embedded as a scalable distribution alignment penalty in the loss function, enabling robust federated validation and training. Experiments demonstrate that FIRE achieves up to 5.1% higher accuracy than importance-weighted baselines on shifted validation sets and outperforms standard federated learning methods by 5.3%, significantly enhancing cross-distribution generalization.

0 citationsRead paper

Confidence-Calibrating Regularization for Robust Brain MRI Segmentation Under Domain Shift

Sep 27, 2025

To address poor generalization and inadequate uncertainty calibration of the Segment Anything Model (SAM) in brain MRI segmentation—caused by domain shift and overconfidence—this paper proposes CalSAM, a lightweight adaptive framework. CalSAM freezes SAM’s encoder and fine-tunes only the decoder. It introduces two novel penalties: (i) a feature-level Fisher information penalty to suppress domain sensitivity, and (ii) a voxel-wise confidence misalignment penalty to calibrate prediction reliability. This dual-penalty mechanism jointly enhances robustness and calibration across multi-center and multi-device scenarios while preserving computational efficiency. On the BraTS scanner transfer task, CalSAM achieves a Dice Similarity Coefficient (DSC) of 80.1% (+significant gain) and reduces Hausdorff Distance at 95% (HD95) by 26.9%. On ATLAS-C motion-corrupted data, it attains a DSC of 75.9% and reduces Expected Calibration Error (ECE) by 32.6%, consistently outperforming baselines. These results validate CalSAM’s effectiveness and strong cross-domain generalizability.

0 citationsRead paper

Recent publications

Latest Papers

CalTwin: Towards Calibrated, Shift-Robust Medical World Models via Fisher-Information Regularisation

Jul 29, 2026

This work addresses covariate shift and confidence miscalibration in clinical deployment of medical world models by proposing CalTwin, a lightweight regularization method that uniquely integrates Fisher Information Matrix (FIM) penalization with confidence miscalibration penalization into a GRU-based latent state transition predictor. Through joint optimization, CalTwin simultaneously mitigates performance degradation under out-of-distribution data and overconfidence. Evaluated on the PhysioNet 2019 Sepsis Challenge dataset, CalTwin reduces the mean squared error of out-of-distribution next-step latent state prediction by 9.1%—with 7.0% attributable to FIM regularization—and modestly improves calibration, lowering expected calibration error (ECE) by 0.7%, thereby significantly enhancing model reliability and calibration in multi-center, heterogeneous clinical settings.

0 citationsRead paper

PIcsC: Partitioning-Induced Covariate Shift Correction

Jul 28, 2026

This work addresses covariate shift induced by data partitioning, which severely undermines model selection and evaluation in cross-validation, continual learning, and federated learning. The authors propose the PIcsC framework, which introduces a Fisher information–based regularization term to uniformly mitigate partition-induced shifts in both centralized and distributed settings. PIcsC approximates distributional discrepancies—detected via KL divergence—using local gradient statistics and activates conditional adaptive regularization only when significant shifts are identified, without requiring access to raw data. Evaluated across more than 40 datasets, PIcsC substantially alleviates performance degradation, reducing loss by 20% under batch partitioning and by 25% in cross-validation. Moreover, it outperforms baselines such as FedAvg by 3–5 percentage points on seven federated learning benchmarks.

0 citationsRead paper

Technical note on Sequential Test-Time Adaptation via Martingale-Driven Fisher Prompting

Oct 04, 2025

This paper addresses robust sequential decision-making under covariate shift in streaming data. Methodologically, it proposes the M-FISHER framework, which unifies real-time distributional shift detection and stable online adaptation. It constructs an exponential martingale from non-conformity scores and leverages Ville’s inequality to guarantee bounded false alarm rate at any time; additionally, it employs Fisher-preconditioned parameter updates to perform natural gradient descent on the distribution manifold, ensuring geometric invariance and KL-divergence minimization. Theoretically, it is the first to integrate martingale theory with Fisher information geometry for test-time adaptation, delivering time-uniform statistical guarantees; it also derives an explicit upper bound on detection delay—O(log(1/δ)/Γ)—that scales inversely with the magnitude Γ of distributional discrepancy. Experiments demonstrate its superior detection efficiency and adaptive stability.

0 citationsRead paper

Technical note on Fisher Information for Robust Federated Cross-Validation

Oct 04, 2025

In federated learning, spatiotemporally distributed data induce covariate shift, causing local empirical distributions across clients to deviate from the global underlying distribution and thereby degrading model generalization. To address this, we propose FIRE (Fisher Information Regularized Estimation), the first method to incorporate the Fisher information matrix into federated cross-validation. FIRE approximates the Fisher information distance between each client’s local data and the global distribution, quantifying and correcting covariate shift. This distance is embedded as a scalable distribution alignment penalty in the loss function, enabling robust federated validation and training. Experiments demonstrate that FIRE achieves up to 5.1% higher accuracy than importance-weighted baselines on shifted validation sets and outperforms standard federated learning methods by 5.3%, significantly enhancing cross-distribution generalization.

0 citationsRead paper

Confidence-Calibrating Regularization for Robust Brain MRI Segmentation Under Domain Shift

Sep 27, 2025

To address poor generalization and inadequate uncertainty calibration of the Segment Anything Model (SAM) in brain MRI segmentation—caused by domain shift and overconfidence—this paper proposes CalSAM, a lightweight adaptive framework. CalSAM freezes SAM’s encoder and fine-tunes only the decoder. It introduces two novel penalties: (i) a feature-level Fisher information penalty to suppress domain sensitivity, and (ii) a voxel-wise confidence misalignment penalty to calibrate prediction reliability. This dual-penalty mechanism jointly enhances robustness and calibration across multi-center and multi-device scenarios while preserving computational efficiency. On the BraTS scanner transfer task, CalSAM achieves a Dice Similarity Coefficient (DSC) of 80.1% (+significant gain) and reduces Hausdorff Distance at 95% (HD95) by 26.9%. On ATLAS-C motion-corrupted data, it attains a DSC of 75.9% and reduces Expected Calibration Error (ECE) by 32.6%, consistently outperforming baselines. These results validate CalSAM’s effectiveness and strong cross-domain generalizability.

0 citationsRead paper