Zeyu Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Zeyu Li has a total of 35 co-authors. Here are the top co-authors by number of shared publications:
Qiang Yan
5 shared publicationsWeipeng Kong
5 shared publicationsMingrui Zou
Research Center of Laser Fusion, China Academy of Engineering Physics , Mianyang 621900, China
4 shared publicationsYu Qin
Research Center of Laser Fusion, China Academy of Engineering Physics , Mianyang 621900, China
4 shared publicationsXun Zhou
3 shared publications- Xun Zhou
Research Center of Laser Fusion, China Academy of Engineering Physics, Mianyang, China
3 shared publications
Latest Publications
- Open Access
Variational Trajectory Optimization of Anisotropic Diffusion Schedules
We introduce a variational framework for diffusion models with anisotropic noise schedules parameterized by a matrix-valued path $M_t(θ)$ that allocates noise across subspaces. Central to our framework is a trajectory-level objective that jointly trains the score network and learns $M_t(θ)$, which encompasses general parameterization classes of matrix-valued noise schedules. We further derive an estimator for the derivative with respect to $θ$ of the score that enables efficient optimization of the $M_t(θ)$ schedule. For inference, we develop an efficiently-implementable reverse-ODE solver that is an anisotropic generalization of the second-order Heun discretization algorithm. Across CIFAR-10, AFHQv2, FFHQ, and ImageNet-64, our method consistently improves upon the baseline EDM model in all NFE regimes. Code is available at https://github.com/lizeyu090312/anisotropic-diffusion-paper.Open AccessVariational Trajectory Optimization of Anisotropic Diffusion Schedules
We introduce a variational framework for diffusion models with anisotropic noise schedules parameterized by a matrix-valued path $M_t(θ)$ that allocates noise across subspaces. Central to our framework is a trajectory-level objective that jointly trains the score network and learns $M_t(θ)$, which encompasses general parameterization classes of matrix-valued noise schedules. We further derive an estimator for the derivative with respect to $θ$ of the score that enables efficient optimization of the $M_t(θ)$ schedule. For inference, we develop an efficiently-implementable reverse-ODE solver that is an anisotropic generalization of the second-order Heun discretization algorithm. Across CIFAR-10, AFHQv2, FFHQ, and ImageNet-64, our method consistently improves upon the baseline EDM model in all NFE regimes. Code is available at https://github.com/lizeyu090312/anisotropic-diffusion-paper. - 19 December 2024
- Published by Elsevier in Infrared Physics & Technology
- Vol. 145
Dual-plane synthetic aperture phase retrieval with the CAEP pulsed terahertz free electron laser
- 19 December 2024
- Published by Elsevier in Infrared Physics & Technology
- Vol. 145
Dual-plane synthetic aperture phase retrieval with the CAEP pulsed terahertz free electron laser
- Open Access
Using Interleaved Ensemble Unlearning to Keep Backdoors at Bay for Finetuning Vision Transformers
Vision Transformers (ViTs) have become popular in computer vision tasks. Backdoor attacks, which trigger undesirable behaviours in models during inference, threaten ViTs' performance, particularly in security-sensitive tasks. Although backdoor defences have been developed for Convolutional Neural Networks (CNNs), they are less effective for ViTs, and defences tailored to ViTs are scarce. To address this, we present Interleaved Ensemble Unlearning (IEU), a method for finetuning clean ViTs on backdoored datasets. In stage 1, a shallow ViT is finetuned to have high confidence on backdoored data and low confidence on clean data. In stage 2, the shallow ViT acts as a ``gate'' to block potentially poisoned data from the defended ViT. This data is added to an unlearn set and asynchronously unlearned via gradient ascent. We demonstrate IEU's effectiveness on three datasets against 11 state-of-the-art backdoor attacks and show its versatility by applying it to different model architectures.
Open AccessUsing Interleaved Ensemble Unlearning to Keep Backdoors at Bay for Finetuning Vision Transformers
Vision Transformers (ViTs) have become popular in computer vision tasks. Backdoor attacks, which trigger undesirable behaviours in models during inference, threaten ViTs' performance, particularly in security-sensitive tasks. Although backdoor defences have been developed for Convolutional Neural Networks (CNNs), they are less effective for ViTs, and defences tailored to ViTs are scarce. To address this, we present Interleaved Ensemble Unlearning (IEU), a method for finetuning clean ViTs on backdoored datasets. In stage 1, a shallow ViT is finetuned to have high confidence on backdoored data and low confidence on clean data. In stage 2, the shallow ViT acts as a ``gate'' to block potentially poisoned data from the defended ViT. This data is added to an unlearn set and asynchronously unlearned via gradient ascent. We demonstrate IEU's effectiveness on three datasets against 11 state-of-the-art backdoor attacks and show its versatility by applying it to different model architectures.
- 1 April 2024
- Published by Elsevier in Engineering Structures
- Vol. 307, 117911
Nonlinear static solution of suspension bridge formulated from a surrogate model with secant stiffness of suspension cables
- 1 April 2024
- Published by Elsevier in Engineering Structures
- Vol. 307, 117911
Nonlinear static solution of suspension bridge formulated from a surrogate model with secant stiffness of suspension cables
- Open Access
Geo2SigMap: High-Fidelity RF Signal Mapping Using Geographic Databases
Radio frequency (RF) signal mapping, which is the process of analyzing and predicting the RF signal strength and distribution across specific areas, is crucial for cellular network planning and deployment. Traditional approaches to RF signal mapping rely on statistical models constructed based on measurement data, which offer low complexity but often lack accuracy, or ray tracing tools, which provide enhanced precision for the target area but suffer from increased computational complexity. Recently, machine learning (ML) has emerged as a data-driven method for modeling RF signal propagation, which leverages models trained on synthetic datasets to perform RF signal mapping in "unseen" areas. In this paper, we present Geo2SigMap, an ML-based framework for efficient and high-fidelity RF signal mapping using geographic databases. First, we develop an automated framework that seamlessly integrates three open-source tools: OpenStreetMap (geographic databases), Blender (computer graphics), and Sionna (ray tracing), enabling the efficient generation of large-scale 3D building maps and ray tracing models. Second, we propose a cascaded U-Net model, which is pre-trained on synthetic datasets and employed to generate detailed RF signal maps, leveraging environmental information and sparse measurement data. Finally, we evaluate the performance of Geo2SigMap via a real-world measurement campaign, where three types of user equipment (UE) collect over 45,000 data points related to cellular information from six LTE cells operating in the citizens broadband radio service (CBRS) band. Our results show that Geo2SigMap achieves an average root-mean-square-error (RMSE) of 6.04 dB for predicting the reference signal received power (RSRP) at the UE, representing an average RMSE improvement of 3.59 dB compared to existing methods.
Open AccessGeo2SigMap: High-Fidelity RF Signal Mapping Using Geographic Databases
Radio frequency (RF) signal mapping, which is the process of analyzing and predicting the RF signal strength and distribution across specific areas, is crucial for cellular network planning and deployment. Traditional approaches to RF signal mapping rely on statistical models constructed based on measurement data, which offer low complexity but often lack accuracy, or ray tracing tools, which provide enhanced precision for the target area but suffer from increased computational complexity. Recently, machine learning (ML) has emerged as a data-driven method for modeling RF signal propagation, which leverages models trained on synthetic datasets to perform RF signal mapping in "unseen" areas. In this paper, we present Geo2SigMap, an ML-based framework for efficient and high-fidelity RF signal mapping using geographic databases. First, we develop an automated framework that seamlessly integrates three open-source tools: OpenStreetMap (geographic databases), Blender (computer graphics), and Sionna (ray tracing), enabling the efficient generation of large-scale 3D building maps and ray tracing models. Second, we propose a cascaded U-Net model, which is pre-trained on synthetic datasets and employed to generate detailed RF signal maps, leveraging environmental information and sparse measurement data. Finally, we evaluate the performance of Geo2SigMap via a real-world measurement campaign, where three types of user equipment (UE) collect over 45,000 data points related to cellular information from six LTE cells operating in the citizens broadband radio service (CBRS) band. Our results show that Geo2SigMap achieves an average root-mean-square-error (RMSE) of 6.04 dB for predicting the reference signal received power (RSRP) at the UE, representing an average RMSE improvement of 3.59 dB compared to existing methods.
- Open Access
- 1 November 2022
- Published by AIP Publishing in APL Photonics
- Vol. 7 (11) , 116102
Subwavelength full-field terahertz ptychography via longitudinal shifts
We demonstrate subwavelength resolution, full-field, and lensless terahertz ptychography via longitudinal shifts of the sample. An extended illumination is adopted to realize full-field recording. Compared to a typical scheme, this implementation requires fewer diffraction patterns, which improves the imaging throughput. To suppress the incoherent infrared background, an optical chopper is introduced and a preprocessing algorithm based on digital phase-locked is proposed. By combining the ideas of multi-plane phase retrieval and ptychographic reconstruction, a reconstruction algorithm is proposed to simultaneously retrieve the object and the probe. Due to the lensless single-beam mode and full-field illumination, the sample is placed as close to the detector as possible, and the aperture of the whole sensor is utilized to achieve the highest numerical aperture. Note that the proposed approach could be applied to other wavebands as well. We apply this method to a biological material, a polymer sample, and a resolution target. The uneven illumination, as well as the diffraction effects associated with detector edges, is separated from the object reconstruction. The secondary veins of the dragonfly forewing, as small as ∼20 µm, are observed, and the lateral resolution achieves 88 µm (0.74λ).
- 1 November 2022
- Published by AIP Publishing in APL Photonics
- Vol. 7 (11) , 116102
Open AccessSubwavelength full-field terahertz ptychography via longitudinal shifts
We demonstrate subwavelength resolution, full-field, and lensless terahertz ptychography via longitudinal shifts of the sample. An extended illumination is adopted to realize full-field recording. Compared to a typical scheme, this implementation requires fewer diffraction patterns, which improves the imaging throughput. To suppress the incoherent infrared background, an optical chopper is introduced and a preprocessing algorithm based on digital phase-locked is proposed. By combining the ideas of multi-plane phase retrieval and ptychographic reconstruction, a reconstruction algorithm is proposed to simultaneously retrieve the object and the probe. Due to the lensless single-beam mode and full-field illumination, the sample is placed as close to the detector as possible, and the aperture of the whole sensor is utilized to achieve the highest numerical aperture. Note that the proposed approach could be applied to other wavebands as well. We apply this method to a biological material, a polymer sample, and a resolution target. The uneven illumination, as well as the diffraction effects associated with detector edges, is separated from the object reconstruction. The secondary veins of the dragonfly forewing, as small as ∼20 µm, are observed, and the lateral resolution achieves 88 µm (0.74λ).
- 14 September 2022
- Published by Elsevier in Materials Letters
- Vol. 328, 133183
A facile one-step gel process for precipitation of multifractal NaCl crystals and morphology evolution from sodium silicate gel
- 14 September 2022
- Published by Elsevier in Materials Letters
- Vol. 328, 133183
A facile one-step gel process for precipitation of multifractal NaCl crystals and morphology evolution from sodium silicate gel
- 18 May 2021
- Published by Springer Nature in Journal of Infrared, Millimeter, and Terahertz Waves
- Vol. 42 (5) , 479-492
Resolution Enhancement in Terahertz Digital In-line Holography by Sparsity-Based Extrapolation
- 18 May 2021
- Published by Springer Nature in Journal of Infrared, Millimeter, and Terahertz Waves
- Vol. 42 (5) , 479-492
Resolution Enhancement in Terahertz Digital In-line Holography by Sparsity-Based Extrapolation
- 25 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Terahertz Science and Technology
- Vol. 11 (4) , 389-395
Nonlinear Optical Up-Conversion in Terahertz Parametric Process: Simulation and Experiment
We present a numerical method to model the nonlinear optical up-conversion in the terahertz (THz) parametric process. Specifically, we introduce the concept of competition between the spontaneous parametric down-conversion (SPDC) idler beam and the up-converted Stokes beam. We simulate the spectra with the absolute power of each frequency component that clearly shows the competition result. By this method, we theoretically explain the suppression effect of injected signal on the SPDC idler background. We also verify the simulation experimentally by using an up-conversion parametric detection system. We show the competition result between the Stokes signal and SPDC idler background, the different levels of suppression caused by different input THz energy. The actual detection ability of the up-conversion system is also evaluated. All these experimental results are in good agreement with theoretical simulations. This research enables a comprehensive understanding of the physics of the up-conversion detection as well as the THz parametric generation and the seed injection.
- 25 January 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Terahertz Science and Technology
- Vol. 11 (4) , 389-395
Nonlinear Optical Up-Conversion in Terahertz Parametric Process: Simulation and Experiment
We present a numerical method to model the nonlinear optical up-conversion in the terahertz (THz) parametric process. Specifically, we introduce the concept of competition between the spontaneous parametric down-conversion (SPDC) idler beam and the up-converted Stokes beam. We simulate the spectra with the absolute power of each frequency component that clearly shows the competition result. By this method, we theoretically explain the suppression effect of injected signal on the SPDC idler background. We also verify the simulation experimentally by using an up-conversion parametric detection system. We show the competition result between the Stokes signal and SPDC idler background, the different levels of suppression caused by different input THz energy. The actual detection ability of the up-conversion system is also evaluated. All these experimental results are in good agreement with theoretical simulations. This research enables a comprehensive understanding of the physics of the up-conversion detection as well as the THz parametric generation and the seed injection.
- 15 November 2019
- Published by Optica Publishing Group in Photonics Research
- Vol. 7 (12) , 1391-1399
Terahertz synthetic aperture in-line holography with intensity correction and sparsity autofocusing reconstruction
We demonstrate high-resolution and high-quality terahertz (THz) in-line digital holography based on the synthetic aperture method. The setup is built on a self-developed THz quantum cascade laser, and a lateral resolution better than 70 μm () is achieved at 4.3 THz. To correct intensity differences between sub-holograms before aperture stitching, a practical algorithm with global optimization is proposed. To address the twin-image problem for in-line holography, a sparsity-based phase retrieval algorithm is applied to perform the high-quality reconstruction. Furthermore, a new autofocusing criterion termed “reconstruction objective function” is introduced to obtain the best in-focus reconstruction distance, so the autofocusing procedure and the reconstruction are unified within the same framework. Both simulation and experiment prove its accuracy and robustness. Note that all the methods proposed here can be applied to other wavebands as well. We demonstrate the success of this THz synthetic aperture in-line holography on biological and semiconductor samples, showing its potential applications in bioimaging and materials analysis.- 15 November 2019
- Published by Optica Publishing Group in Photonics Research
- Vol. 7 (12) , 1391-1399
Terahertz synthetic aperture in-line holography with intensity correction and sparsity autofocusing reconstruction
We demonstrate high-resolution and high-quality terahertz (THz) in-line digital holography based on the synthetic aperture method. The setup is built on a self-developed THz quantum cascade laser, and a lateral resolution better than 70 μm () is achieved at 4.3 THz. To correct intensity differences between sub-holograms before aperture stitching, a practical algorithm with global optimization is proposed. To address the twin-image problem for in-line holography, a sparsity-based phase retrieval algorithm is applied to perform the high-quality reconstruction. Furthermore, a new autofocusing criterion termed “reconstruction objective function” is introduced to obtain the best in-focus reconstruction distance, so the autofocusing procedure and the reconstruction are unified within the same framework. Both simulation and experiment prove its accuracy and robustness. Note that all the methods proposed here can be applied to other wavebands as well. We demonstrate the success of this THz synthetic aperture in-line holography on biological and semiconductor samples, showing its potential applications in bioimaging and materials analysis.