Selected work

Representative Papers

Neural Variability Enhances Artificial Network Robustness

Jun 11, 2026

This work addresses the limited robustness of artificial neural networks under adversarial attacks and natural image corruptions, a challenge often exacerbated by the neglect of structured noise in neural activations. The authors propose a biologically inspired local noise mechanism that models structured noise by analyzing the covariance structure of activations induced by clean and perturbed inputs. Relying solely on local information, this approach is the first to systematically reveal how structured noise differentially enhances robustness across perturbation types. Experimental results demonstrate that the proposed strategy significantly improves model robustness against natural corruptions, and notably, the noise structures learned under adversarial attacks exhibit strong generalization to other attack variants.

0 citationsRead paper

A Biconvex Formulation for Stable Transport of Mixture Models with a Unique Solution

Jun 01, 2026

Traditional optimal transport suffers from high computational costs and limited interpretability when applied to large-scale data. This work proposes an Optimal Mixture Transport (OMT) framework that elevates the transport unit from individual samples to subpopulation-level mixture models. By modeling subpopulations with exponential family distributions, the problem is reformulated as a strictly biconvex optimization, yielding—for the first time—a mixture transport method with guarantees of a unique global solution and stability. Notably, the computational complexity of OMT depends only on the number of mixture components and is decoupled from the sample size. Experiments demonstrate that OMT achieves superior efficiency, stability, and interpretability across synthetic data, image tasks, and large-scale single-cell RNA sequencing applications.

0 citationsRead paper

A Fokker-Planck-Based Loss Function that Bridges Dynamics with Density Estimation

Feb 24, 2025

This work establishes a theoretical link between dynamical system models and their stationary probability density functions (PDFs), enabling dynamics parameter inference without time-series data and high-fidelity PDF estimation under known dynamics. To this end, we propose a novel physics-informed loss function grounded in the Fokker–Planck equation—constituting the first unified framework jointly optimizing dynamical modeling and density estimation. We further design a hybrid density estimator integrating Gaussian Mixture Models (GMMs) with normalizing flows, augmented by a Hopfield-like energy-based latent space; optimization employs the Concave-Convex Procedure (CCCP) for interpretable and efficient data manipulation. Experiments on noisy Lorenz systems and gene regulatory networks demonstrate: (i) timestamp-free parameter identification; (ii) significantly improved PDF estimation accuracy in sparse regions; and (iii) effective support for downstream tasks including denoising and clustering—achieving both theoretical rigor and practical utility.

0 citationsRead paper

Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function

Jan 02, 2025

In neural connectomics, accurate axon reconstruction is hindered by topological errors—such as fragmentation and merging—in instance segmentation of highly entangled filamentous structures. To address this, we propose a lightweight topology-aware segmentation method. Our approach introduces, for the first time, the digital topology concept of “simple points” to supervoxel-level connected components, enabling a supervoxel-based topological constraint loss that preserves connectivity while maintaining computational efficiency. Integrated within a 3D U-Net architecture, this loss is coupled with digital-topology-guided connectivity regularization. Evaluated on a novel mouse brain light-sheet microscopy dataset and established benchmarks (DRIVE, ISBI12, CrackTree), our method significantly reduces fragmentation and merging errors, improves instance-level connectivity accuracy, and incurs negligible computational overhead. The key contributions are: (i) the generalization of simple-point theory to supervoxels; (ii) a differentiable, topology-preserving loss; and (iii) a computationally efficient framework achieving state-of-the-art topological fidelity in filamentous structure segmentation.

0 citationsRead paper

Recent publications

Latest Papers

Neural Variability Enhances Artificial Network Robustness

Jun 11, 2026

This work addresses the limited robustness of artificial neural networks under adversarial attacks and natural image corruptions, a challenge often exacerbated by the neglect of structured noise in neural activations. The authors propose a biologically inspired local noise mechanism that models structured noise by analyzing the covariance structure of activations induced by clean and perturbed inputs. Relying solely on local information, this approach is the first to systematically reveal how structured noise differentially enhances robustness across perturbation types. Experimental results demonstrate that the proposed strategy significantly improves model robustness against natural corruptions, and notably, the noise structures learned under adversarial attacks exhibit strong generalization to other attack variants.

0 citationsRead paper

A Biconvex Formulation for Stable Transport of Mixture Models with a Unique Solution

Jun 01, 2026

Traditional optimal transport suffers from high computational costs and limited interpretability when applied to large-scale data. This work proposes an Optimal Mixture Transport (OMT) framework that elevates the transport unit from individual samples to subpopulation-level mixture models. By modeling subpopulations with exponential family distributions, the problem is reformulated as a strictly biconvex optimization, yielding—for the first time—a mixture transport method with guarantees of a unique global solution and stability. Notably, the computational complexity of OMT depends only on the number of mixture components and is decoupled from the sample size. Experiments demonstrate that OMT achieves superior efficiency, stability, and interpretability across synthetic data, image tasks, and large-scale single-cell RNA sequencing applications.

0 citationsRead paper

A Fokker-Planck-Based Loss Function that Bridges Dynamics with Density Estimation

Feb 24, 2025

This work establishes a theoretical link between dynamical system models and their stationary probability density functions (PDFs), enabling dynamics parameter inference without time-series data and high-fidelity PDF estimation under known dynamics. To this end, we propose a novel physics-informed loss function grounded in the Fokker–Planck equation—constituting the first unified framework jointly optimizing dynamical modeling and density estimation. We further design a hybrid density estimator integrating Gaussian Mixture Models (GMMs) with normalizing flows, augmented by a Hopfield-like energy-based latent space; optimization employs the Concave-Convex Procedure (CCCP) for interpretable and efficient data manipulation. Experiments on noisy Lorenz systems and gene regulatory networks demonstrate: (i) timestamp-free parameter identification; (ii) significantly improved PDF estimation accuracy in sparse regions; and (iii) effective support for downstream tasks including denoising and clustering—achieving both theoretical rigor and practical utility.

0 citationsRead paper

Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function

Jan 02, 2025

In neural connectomics, accurate axon reconstruction is hindered by topological errors—such as fragmentation and merging—in instance segmentation of highly entangled filamentous structures. To address this, we propose a lightweight topology-aware segmentation method. Our approach introduces, for the first time, the digital topology concept of “simple points” to supervoxel-level connected components, enabling a supervoxel-based topological constraint loss that preserves connectivity while maintaining computational efficiency. Integrated within a 3D U-Net architecture, this loss is coupled with digital-topology-guided connectivity regularization. Evaluated on a novel mouse brain light-sheet microscopy dataset and established benchmarks (DRIVE, ISBI12, CrackTree), our method significantly reduces fragmentation and merging errors, improves instance-level connectivity accuracy, and incurs negligible computational overhead. The key contributions are: (i) the generalization of simple-point theory to supervoxels; (ii) a differentiable, topology-preserving loss; and (iii) a computationally efficient framework achieving state-of-the-art topological fidelity in filamentous structure segmentation.

0 citationsRead paper