Selected work

Representative Papers

CBIL: Collective Behavior Imitation Learning for Fish from Real Videos

Nov 19, 2024ACM Transactions on Graphics

This work addresses the challenge of modeling collective behaviors in high-density, irregular fish schools. We propose the first end-to-end, video-driven self-supervised imitation learning framework that learns spatiotemporal motion patterns directly from raw videos—without requiring ground-truth trajectory annotations. Methodologically, the framework integrates a Masked Video Autoencoder (MVAE) with self-supervised video representation learning to extract robust spatiotemporal features; introduces an adversarial latent-space motion distribution matching mechanism; and incorporates biologically inspired reward functions and prior motion constraints to enhance training stability and behavioral plausibility. Experiments demonstrate substantial improvements in behavioral diversity and visual fidelity of generated motions. The framework generalizes effectively to multi-species animated synthesis and enables automatic detection of anomalous schooling behaviors in field-captured videos.

2 citationsRead paper

From Rubble Simulation to Active Magnetic Mapping: Quantum Sensing for Disaster Response

Jun 24, 2026

This study addresses the challenge of accurately locating survivors within the critical 72-hour window following building collapse, where limited knowledge of internal rubble structure impedes rescue efforts. The authors propose an active sensing approach employing a drone-mounted array of quantum magnetometers, integrating quantum-grade magnetic sensing with Bayesian active learning for the first time. They develop an end-to-end simulation framework that encompasses physical collapse modeling (based on Unreal Engine), dipole magnetic field approximation via triangular surface elements, and Gaussian process regression–driven spatial magnetic field reconstruction. Experimental results demonstrate effective recovery of magnetic signals ranging from sub-picotesla to sub-nanotesla levels at approximately one meter above the rubble surface. A three-sensor array achieves optimal structural correlation within fewer than 100 sampling iterations, validating the method’s feasibility and efficiency for void detection in disaster scenarios.

0 citationsRead paper

Hodge Spectral Surrogates for Topology-Constrained Optimization

Jun 23, 2026

This work addresses the challenge of directly controlling Betti numbers and persistent homology in optimization or generation tasks due to their discrete and combinatorial nature. It proposes the first framework that integrates Hodge spectral theory with differentiable optimization, constructing a spectral relaxation of the Hodge–Laplacian via soft graphs and soft clique complexes. By introducing spectral proxies for zero and near-zero modes, the method enables smooth, geometry-aware gradient updates toward topological objectives. The approach unifies treatment of Vietoris–Rips filtrations and graph clique complexes, yielding more uniform spatial gradients and smoother scale normalization in point cloud tasks, while effectively regulating normalized first Betti numbers on graphs and supporting joint optimization with standard graph features.

0 citationsRead paper

EML-CD: Causal Mechanism Recovery via EML Symbolic Trees in Structure Learning

Jun 04, 2026

This work addresses the limited interpretability of existing neural network–based causal discovery methods, which recover directed acyclic graphs but treat causal mechanisms as black boxes. To overcome this, the authors propose embedding Elementary Mathematical Language (EML) operators—comprising elementary functions combined via a single binary operator—into a structure learning framework. By employing gated EML symbolic trees, the method enables end-to-end, closed-form discovery of causal equations. It is the first approach to jointly optimize both graph structure and interpretable causal mechanisms within causal discovery, while also supporting analytical computation of Jacobian matrices for quantifying causal effects. Experiments demonstrate that on the Sachs dataset, the method achieves structural recovery performance comparable to PC and GES (SHD = 11.2 ± 0.4) and attains a per-edge closed-form equation accuracy of 0.756. On bivariate benchmarks, it accurately recovers 10 out of 11 elementary function types, with mechanism prediction errors significantly lower than those of SINDy.

0 citationsRead paper

AUGUSTE: Online-Learning dApp for Predictive URLLC Scheduling

Jun 02, 2026

This study addresses the challenge of excessive uplink latency (50–70 ms) in 5G Ultra-Reliable Low-Latency Communication (URLLC) applications caused by conventional Scheduling Request (SR) mechanisms, which fails to meet stringent 1–10 ms requirements. To overcome this limitation, the authors propose AUGUSTE, a novel framework that integrates an online machine learning model at the MAC layer to predict packet arrival times and employs an adaptive state machine to dynamically switch between learning and scheduling modes, enabling on-demand resource pre-allocation. This approach uniquely combines online learning with an adaptive state mechanism for URLLC scheduling, efficiently supporting aperiodic traffic without requiring cross-layer synchronization. Real-world evaluations on OpenAirInterface demonstrate that AUGUSTE reduces median round-trip latency to approximately 10 ms—halving the baseline SR latency—while incurring only 7–10% of the resource overhead associated with persistent scheduling.

0 citationsRead paper

Recent publications

Latest Papers

From Rubble Simulation to Active Magnetic Mapping: Quantum Sensing for Disaster Response

Jun 24, 2026

This study addresses the challenge of accurately locating survivors within the critical 72-hour window following building collapse, where limited knowledge of internal rubble structure impedes rescue efforts. The authors propose an active sensing approach employing a drone-mounted array of quantum magnetometers, integrating quantum-grade magnetic sensing with Bayesian active learning for the first time. They develop an end-to-end simulation framework that encompasses physical collapse modeling (based on Unreal Engine), dipole magnetic field approximation via triangular surface elements, and Gaussian process regression–driven spatial magnetic field reconstruction. Experimental results demonstrate effective recovery of magnetic signals ranging from sub-picotesla to sub-nanotesla levels at approximately one meter above the rubble surface. A three-sensor array achieves optimal structural correlation within fewer than 100 sampling iterations, validating the method’s feasibility and efficiency for void detection in disaster scenarios.

0 citationsRead paper

Hodge Spectral Surrogates for Topology-Constrained Optimization

Jun 23, 2026

This work addresses the challenge of directly controlling Betti numbers and persistent homology in optimization or generation tasks due to their discrete and combinatorial nature. It proposes the first framework that integrates Hodge spectral theory with differentiable optimization, constructing a spectral relaxation of the Hodge–Laplacian via soft graphs and soft clique complexes. By introducing spectral proxies for zero and near-zero modes, the method enables smooth, geometry-aware gradient updates toward topological objectives. The approach unifies treatment of Vietoris–Rips filtrations and graph clique complexes, yielding more uniform spatial gradients and smoother scale normalization in point cloud tasks, while effectively regulating normalized first Betti numbers on graphs and supporting joint optimization with standard graph features.

0 citationsRead paper

EML-CD: Causal Mechanism Recovery via EML Symbolic Trees in Structure Learning

Jun 04, 2026

This work addresses the limited interpretability of existing neural network–based causal discovery methods, which recover directed acyclic graphs but treat causal mechanisms as black boxes. To overcome this, the authors propose embedding Elementary Mathematical Language (EML) operators—comprising elementary functions combined via a single binary operator—into a structure learning framework. By employing gated EML symbolic trees, the method enables end-to-end, closed-form discovery of causal equations. It is the first approach to jointly optimize both graph structure and interpretable causal mechanisms within causal discovery, while also supporting analytical computation of Jacobian matrices for quantifying causal effects. Experiments demonstrate that on the Sachs dataset, the method achieves structural recovery performance comparable to PC and GES (SHD = 11.2 ± 0.4) and attains a per-edge closed-form equation accuracy of 0.756. On bivariate benchmarks, it accurately recovers 10 out of 11 elementary function types, with mechanism prediction errors significantly lower than those of SINDy.

0 citationsRead paper

AUGUSTE: Online-Learning dApp for Predictive URLLC Scheduling

Jun 02, 2026

This study addresses the challenge of excessive uplink latency (50–70 ms) in 5G Ultra-Reliable Low-Latency Communication (URLLC) applications caused by conventional Scheduling Request (SR) mechanisms, which fails to meet stringent 1–10 ms requirements. To overcome this limitation, the authors propose AUGUSTE, a novel framework that integrates an online machine learning model at the MAC layer to predict packet arrival times and employs an adaptive state machine to dynamically switch between learning and scheduling modes, enabling on-demand resource pre-allocation. This approach uniquely combines online learning with an adaptive state mechanism for URLLC scheduling, efficiently supporting aperiodic traffic without requiring cross-layer synchronization. Real-world evaluations on OpenAirInterface demonstrate that AUGUSTE reduces median round-trip latency to approximately 10 ms—halving the baseline SR latency—while incurring only 7–10% of the resource overhead associated with persistent scheduling.

0 citationsRead paper

Gauge Geometry of Hodge Zero-Mode Transport in Parameter-Dependent Topological Data Analysis

May 27, 2026

Existing approaches to persistent homology struggle to characterize the evolution, reorganization, and memory effects of homological features in parameter-dependent topological data. This work proposes a unified framework based on the zero modes of the combinatorial Hodge Laplacian to track homological feature evolution within a shared chain space. For the first time, it incorporates curvature and holonomy from differential geometry to capture local reorganization dynamics and cyclically accumulated memory, respectively. By transcending the representational limitations of traditional persistence diagrams, the method successfully identifies instability in feature tracking in time-varying point cloud experiments, distinguishes systems whose persistence diagrams are highly similar, and reveals higher-order cyclic memory structures that pairwise matching approaches fail to detect.

0 citationsRead paper