Selected work

Representative Papers

3D Point Cloud Object Detection on Edge Devices for Split Computing

May 13, 20242024 IEEE 3rd Real-Time and Intelligent Edge Computing Workshop (RAGE)

To address the high computational overhead, latency, and power consumption of LiDAR point cloud 3D object detection models on edge devices, this work pioneers the application of split computing to point cloud detection. We propose a dual-granularity collaborative inference framework targeting both voxelized representations and internal network features. By partitioning the model either after voxelization or at intermediate layers of the backbone network, our approach achieves balanced edge–cloud workload distribution and enables compression of sensitive features for transmission—preserving detection accuracy while substantially reducing edge-side burden. Experiments demonstrate that post-voxelization splitting reduces total inference time by 70.8% and edge execution time by 90.0%; internal-network splitting achieves up to 57.1% reduction in end-to-end latency and 69.5% reduction in edge computation time. This work establishes a deployable, real-time 3D perception paradigm for resource-constrained edge scenarios.

3 citationsRead paper

Blending Optimal Control and Biologically Plausible Learning for Noise-Robust Physical Neural Networks.

Jan 07, 2025Physical Review Letters

Physical neural networks (PNNs) suffer from prohibitively high training computational costs and poor robustness to measurement noise. To address these challenges, this work proposes a novel model-free training framework that synergistically integrates continuous-time optimal control with direct feedback alignment (DFA)—the first such combination in the literature. Crucially, it eliminates the need for precise physical system modeling while enabling efficient, noise-resilient weight optimization. The method substantially reduces training time and demonstrates superior stability and generalization under strong measurement noise, as validated through both numerical simulations and experiments on a photonic-electronic delay system. It consistently outperforms conventional gradient-based training approaches. This work broadens the class of physically realizable systems suitable for practical PNNs and establishes a new paradigm for robust, hardware-aware neuromorphic computing.

1 citationsRead paper

Photonic reservoir computing with complex networks

Jul 25, 2026

This study systematically investigates the impact of complex network topologies on the performance of large-scale photonic reservoir computing. Using a spatial light modulator, the authors experimentally implement reservoirs with small-world, scale-free, and human connectome-inspired topologies, evaluating them through memory capacity and one-step-ahead prediction of chaotic time series. For the first time, this work provides experimental validation of performance differences across distinct network architectures, demonstrating that small-world networks achieve superior memory capacity and prediction accuracy. Furthermore, the study reveals that the rewiring probability of the small-world topology and the reservoir’s leakage rate can be jointly tuned to significantly enhance predictive performance.

0 citationsRead paper

Approximate reservoir computing with a semiconductor laser for reducing energy consumption

Jul 25, 2026

This work addresses the lack of systematic investigation into the joint impact of quantization bit-width, sampling frequency, and hardware parameters on energy efficiency in photonic reservoir computing. The authors propose an approximate reservoir computing framework based on semiconductor lasers, which achieves a balance between high prediction accuracy and low energy consumption by co-optimizing amplitude quantization of node states and output weights, tunable sampling frequency, and injection current. Evaluated on chaotic time series prediction tasks, the approach maintains excellent performance while significantly reducing energy consumption per sample. This study is the first to systematically elucidate the influence mechanism of key parameters on the energy-performance trade-off and demonstrates the feasibility and advantages of approximate computing in photonic neuromorphic systems.

0 citationsRead paper

Profiling Lightweight Large Language Models

Jul 22, 2026

This work addresses the limitations of existing proxy metrics—such as parameter count and FLOPs—in evaluating lightweight large language models, which often fail to capture the true trade-offs between accuracy and resource consumption in real-world deployments. To overcome this, we propose PTME, a novel framework that enables precision-aware, multidimensional resource profiling through direct hardware measurements on local edge devices. PTME jointly evaluates model performance across code generation, mathematical reasoning, and multitask comprehension in terms of accuracy, execution time, peak memory usage, and energy consumption. Our experiments reveal that static proxy metrics poorly predict actual accuracy, that tightening resource constraints primarily increases latency rather than energy use, and that no single model dominates across all dimensions. By identifying Pareto-optimal solutions, PTME provides a data-driven foundation for deployment optimization.

0 citationsRead paper

Recent publications

Latest Papers

Photonic reservoir computing with complex networks

Jul 25, 2026

This study systematically investigates the impact of complex network topologies on the performance of large-scale photonic reservoir computing. Using a spatial light modulator, the authors experimentally implement reservoirs with small-world, scale-free, and human connectome-inspired topologies, evaluating them through memory capacity and one-step-ahead prediction of chaotic time series. For the first time, this work provides experimental validation of performance differences across distinct network architectures, demonstrating that small-world networks achieve superior memory capacity and prediction accuracy. Furthermore, the study reveals that the rewiring probability of the small-world topology and the reservoir’s leakage rate can be jointly tuned to significantly enhance predictive performance.

0 citationsRead paper

Approximate reservoir computing with a semiconductor laser for reducing energy consumption

Jul 25, 2026

This work addresses the lack of systematic investigation into the joint impact of quantization bit-width, sampling frequency, and hardware parameters on energy efficiency in photonic reservoir computing. The authors propose an approximate reservoir computing framework based on semiconductor lasers, which achieves a balance between high prediction accuracy and low energy consumption by co-optimizing amplitude quantization of node states and output weights, tunable sampling frequency, and injection current. Evaluated on chaotic time series prediction tasks, the approach maintains excellent performance while significantly reducing energy consumption per sample. This study is the first to systematically elucidate the influence mechanism of key parameters on the energy-performance trade-off and demonstrates the feasibility and advantages of approximate computing in photonic neuromorphic systems.

0 citationsRead paper

Profiling Lightweight Large Language Models

Jul 22, 2026

This work addresses the limitations of existing proxy metrics—such as parameter count and FLOPs—in evaluating lightweight large language models, which often fail to capture the true trade-offs between accuracy and resource consumption in real-world deployments. To overcome this, we propose PTME, a novel framework that enables precision-aware, multidimensional resource profiling through direct hardware measurements on local edge devices. PTME jointly evaluates model performance across code generation, mathematical reasoning, and multitask comprehension in terms of accuracy, execution time, peak memory usage, and energy consumption. Our experiments reveal that static proxy metrics poorly predict actual accuracy, that tightening resource constraints primarily increases latency rather than energy use, and that no single model dominates across all dimensions. By identifying Pareto-optimal solutions, PTME provides a data-driven foundation for deployment optimization.

0 citationsRead paper

Toward the Right Analytical Model and System Software for Autonomous Driving Systems: Open Problems and Research Directions

Jul 05, 2026

Autonomous driving systems currently lack temporal analysis models and implementable software that jointly account for multi-rate asynchronous sensor streams and complex actuation chains, hindering guarantees of end-to-end timing correctness and functional safety. This work proposes, for the first time, a five-dimensional research gap framework—encompassing end-to-end latency, data freshness, temporal skew, probabilistic timing, and fail-safe mechanisms—from a co-design perspective of analytical models and system software. By integrating real-time scheduling theories (e.g., DAG-based and mixed-criticality systems), event- and time-triggered paradigms, ROS 2/Autoware architectures, communication optimizations, and runtime tracing techniques, the study systematically uncovers limitations in current approaches regarding constraint modeling, temporal metrics, resource abstraction, execution variability, and safety integration. The findings lay a theoretical and technical foundation for building a highly reliable temporal assurance framework for autonomous vehicles that is analyzable, observable, and deployable.

0 citationsRead paper

VFILC: Accurate Frequency Extrapolations in Imitation Learning via Sampling Frequency ILC

Jun 18, 2026

This work addresses the instability and frequency inaccuracies inherent in existing neural network-based imitation learning when extrapolating to higher movement speeds. While variable-frequency imitation learning (VFIL) suffers from significant frequency errors under high-speed extrapolation due to its open-loop architecture, this study introduces iterative learning control (ILC) into the VFIL framework for the first time, establishing a closed-loop system that combines feedforward and feedback mechanisms. The feedforward component enables frequency-aware speed extrapolation, while the feedback module online corrects frequency deviations. Evaluated across three distinct tasks, the proposed method substantially enhances both frequency accuracy and motion stability: when extrapolating to twice the average training speed, frequency errors are reduced by 81% and 50% in wiping and shaking tasks, respectively; even in interpolation scenarios involving complex friction in a mixed task, accuracy improves by 27%.

0 citationsRead paper