arXiv subjects

Pengfei Lyu

Publications and source records attributed to Pengfei Lyu.

14 recordsLinked to original sources

arXiv2026

When Does Synthetic Data Augmentation Improve Score-Based Imbalanced Classification?

Synthetic data augmentation is widely used to mitigate class imbalance, but its theoretical effects on score-based classification remain poorly understood. This paper develops a framework for characterizing when synthetic minority augmentation can improve threshold-integrated and threshold-optimized metrics, including AUROC, AUPRC, best-threshold balanced accuracy, and best-threshold $\F_1$ score. We separate the effect of augmentation into two components: a change in effective class weighting and a discrepancy between the synthetic and true minority distributions. Under well-specified score models, the raw estimator already targets the likelihood-ratio ordering, which is population-optimal for the metrics considered. Consequently, augmentation cannot provide a fundamental population-level improvement beyond possible finite-sample variance reduction, and may introduce additional bias through synthetic distributional error. We further establish minimax lower bounds showing that the raw estimator already achieves the optimal metric-regret rate in the well-specified regime. Under misspecification, however, augmentation can play a qualitatively different role: by changing the effective class balance, it can alter the restricted-class projection and correct ranking errors induced by the raw imbalanced objective. We provide explicit improvement bounds quantifying the roles of approximation error, finite-sample estimation error, and synthetic distributional error. Simulation studies corroborate the theory, demonstrating limited gains under well-specification and nontrivial but nonmonotone improvements under misspecification.

stat.ML↗

arXiv2025

Bias-Corrected Data Synthesis for Imbalanced Learning

Imbalanced data, where the positive samples represent only a small proportion compared to the negative samples, makes it challenging for classification problems to balance the false positive and false negative rates. A common approach to addressing the challenge involves generating synthetic data for the minority group and then training classification models with both observed and synthetic data. However, since the synthetic data depends on the observed data and fails to replicate the original data distribution accurately, prediction accuracy is reduced when the synthetic data is na\"{i}vely treated as the true data. In this paper, we address the bias introduced by synthetic data and provide consistent estimators for this bias by borrowing information from the majority group. We propose a bias correction procedure to mitigate the adverse effects of synthetic data, enhancing prediction accuracy while avoiding overfitting. This procedure is extended to broader scenarios with imbalanced data, such as imbalanced multi-task learning and causal inference. Theoretical properties, including bounds on bias estimation errors and improvements in prediction accuracy, are provided. Simulation results and data analysis on handwritten digit datasets demonstrate the effectiveness of our method.

stat.ML↗

arXiv2025

Semi-Supervised 3D Medical Segmentation from 2D Natural Images Pretrained Model

This paper explores the transfer of knowledge from general vision models pretrained on 2D natural images to improve 3D medical image segmentation. We focus on the semi-supervised setting, where only a few labeled 3D medical images are available, along with a large set of unlabeled images. To tackle this, we propose a model-agnostic framework that progressively distills knowledge from a 2D pretrained model to a 3D segmentation model trained from scratch. Our approach, M&N, involves iterative co-training of the two models using pseudo-masks generated by each other, along with our proposed learning rate guided sampling that adaptively adjusts the proportion of labeled and unlabeled data in each training batch to align with the models' prediction accuracy and stability, minimizing the adverse effect caused by inaccurate pseudo-masks. Extensive experiments on multiple publicly available datasets demonstrate that M&N achieves state-of-the-art performance, outperforming thirteen existing semi-supervised segmentation approaches under all different settings. Importantly, ablation studies show that M&N remains model-agnostic, allowing seamless integration with different architectures. This ensures its adaptability as more advanced models emerge. The code is available at https://github.com/pakheiyeung/M-N.

cs.CV↗

arXiv2025

Bridging the Inter-Domain Gap through Low-Level Features for Cross-Modal Medical Image Segmentation

This paper addresses the task of cross-modal medical image segmentation by exploring unsupervised domain adaptation (UDA) approaches. We propose a model-agnostic UDA framework, LowBridge, which builds on a simple observation that cross-modal images share some similar low-level features (e.g., edges) as they are depicting the same structures. Specifically, we first train a generative model to recover the source images from their edge features, followed by training a segmentation model on the generated source images, separately. At test time, edge features from the target images are input to the pretrained generative model to generate source-style target domain images, which are then segmented using the pretrained segmentation network. Despite its simplicity, extensive experiments on various publicly available datasets demonstrate that \proposed achieves state-of-the-art performance, outperforming eleven existing UDA approaches under different settings. Notably, further ablation studies show that \proposed is agnostic to different types of generative and segmentation models, suggesting its potential to be seamlessly plugged with the most advanced models to achieve even more outstanding results in the future. The code is available at https://github.com/JoshuaLPF/LowBridge.

eess.IV↗

arXiv2024

Deep Fourier-embedded Network for RGB and Thermal Salient Object Detection

The rapid development of deep learning has significantly improved salient object detection (SOD) combining both RGB and thermal (RGB-T) images. However, existing Transformer-based RGB-T SOD models with quadratic complexity are memory-intensive, limiting their application in high-resolution bimodal feature fusion. To overcome this limitation, we propose a purely Fourier Transform-based model, namely Deep Fourier-embedded Network (FreqSal), for accurate RGB-T SOD. Specifically, we leverage the efficiency of Fast Fourier Transform with linear complexity to design three key components: (1) To fuse RGB and thermal modalities, we propose Modal-coordinated Perception Attention, which aligns and enhances bimodal Fourier representation in multiple dimensions; (2) To clarify object edges and suppress noise, we design Frequency-decomposed Edge-aware Block, which deeply decomposes and filters Fourier components of low-level features; (3) To accurately decode features, we propose Fourier Residual Channel Attention Block, which prioritizes high-frequency information while aligning channel-wise global relationships. Additionally, even when converged, existing deep learning-based SOD models' predictions still exhibit frequency gaps relative to ground-truth. To address this problem, we propose Co-focus Frequency Loss, which dynamically weights hard frequencies during edge frequency reconstruction by cross-referencing bimodal edge information in the Fourier domain. Extensive experiments on ten bimodal SOD benchmark datasets demonstrate that FreqSal outperforms twenty-nine existing state-of-the-art bimodal SOD models. Comprehensive ablation studies further validate the value and effectiveness of our newly proposed components. The code is available at https://github.com/JoshuaLPF/FreqSal.

cs.CV↗

arXiv2024

Efficient Fourier Filtering Network with Contrastive Learning for AAV-based Unaligned Bimodal Salient Object Detection

Autonomous aerial vehicle (AAV)-based bi-modal salient object detection (BSOD) aims to segment salient objects in a scene utilizing complementary cues in unaligned RGB and thermal image pairs. However, the high computational expense of existing AAV-based BSOD models limits their applicability to real-world AAV devices. To address this problem, we propose an efficient Fourier filter network with contrastive learning that achieves both real-time and accurate performance. Specifically, we first design a semantic contrastive alignment loss to align the two modalities at the semantic level, which facilitates mutual refinement in a parameter-free way. Second, inspired by the fast Fourier transform that obtains global relevance in linear complexity, we propose synchronized alignment fusion, which aligns and fuses bi-modal features in the channel and spatial dimensions by a hierarchical filtering mechanism. Our proposed model, AlignSal, reduces the number of parameters by 70.0%, decreases the floating point operations by 49.4%, and increases the inference speed by 152.5% compared to the cutting-edge BSOD model (i.e., MROS). Extensive experiments on the AAV RGB-T 2400 and seven bi-modal dense prediction datasets demonstrate that AlignSal achieves both real-time inference speed and better performance and generalizability compared to nineteen state-of-the-art models across most evaluation metrics. In addition, our ablation studies further verify AlignSal's potential in boosting the performance of existing aligned BSOD models on AAV-based unaligned data. The code is available at: https://github.com/JoshuaLPF/AlignSal.

cs.CV↗

arXiv2024

Testing composite null hypotheses with high-dimensional dependent data: a computationally scalable FDR-controlling procedure

Testing composite null hypotheses is fundamental to many scientific applications, including mediation and replicability analyses, and becomes particularly challenging in high-throughput settings involving tens of thousands of features. Existing high-dimensional composite null hypotheses testing often ignores the dependence structure among features, leading to overly conservative or liberal results. To address this limitation, we develop a four-state hidden Markov model (HMM) for bivariate $p$-value sequences arising from two-study replicability analysis. This model captures local dependence among features and accommodates study-specific heterogeneity. Based on the HMM, we propose a multiple testing procedure that asymptotically controls the false discovery rate (FDR). Extending this framework to more than two studies is computationally intensive, with complexity growing exponentially in the number of studies $n$. To address this scalability issue, we introduce a novel e-value framework that reduces computational complexity to quadratic in $n$, while preserving asymptotic FDR control. Extensive simulations demonstrate that our method achieves higher power than existing approaches at the same FDR levels. When applied to genome-wide association studies (GWAS), the proposed approach identifies replicable SNP-level signals that are not detected at the same significance threshold by competing methods.

stat.ME↗

arXiv2024

Neural-Network-based NLOS Identification in Angular Domain at 60-GHz

This paper introduces an identification method that determines whether a millimeter-wave wireless transmission using directional antennas is being established over a line-of-sight (LOS) or a non-line-of-sight (NLOS) cluster for indoor localization applications. The proposed technique utilizes the channel power angular spectrum that is readily available from a beam training process. In particular, the behavior of five different channel metrics, namely the spatial-domain, time-domain, and frequency-domain channel kurtosis, the mean excess delay, and the RMS delay spread, is analyzed using maximum likelihood ratio and artificial neural network. A noticeable difference between LOS and NLOS clusters is observed and assessed for identification. Hypothesis testing shows errors as low as 0.01-0.02 in simulation and 0.04-0.07 in measurements at 60 GHz.

eess.SP↗

arXiv2022

Scalable lipid droplet microarray fabrication, validation, and screening

High throughput screening of small molecules and natural products is costly, requiring significant amounts of time, reagents, and operating space. Although microarrays have proven effective in the miniaturization of screening for certain biochemical assays, such as nucleic acid hybridization or antibody binding, they are not widely used for drug discovery in cell culture due to the need for cells to internalize lipophilic drug candidates. Lipid droplet microarrays are a promising solution to this problem as they are capable of delivering lipophilic drugs to cells at dosages comparable to solution delivery. However, the scalablility of the array fabrication, assay validation, and screening steps has limited the utility of this approach. Here we demonstrate a scalable process for lipid droplet array fabrication, assay validation in cell culture, and drug screening. A nanointaglio printing process has been adapted for use with a printing press. The arrays are stabilized for immersion into aqueous solution using a vapor coating process. In addition to delivery of lipophilic compounds, we found that we are also able to encapsulate and deliver a water-soluble compound in this way. The arrays can be functionalized by extracellular matrix proteins such as collagen prior to cell culture as the mechanism for uptake is based on direct contact with the lipid delivery vehicles rather than diffusion of the drug out of the microarray spots. We demonstrate this method for delivery to 3 different cell types and the screening of 90 natural product extracts on a microarray covering an area of less than 0.1 cm2. The arrays are suitable for miniaturized screening, for instance in BSL-3 conditions where space is limited and for applications where cell numbers are limited, such as in functional precision medicine.

q-bio.BM↗

arXiv2021

Angular Clustering of Millimeter-Wave Propagation Channels with Watershed Transformation

An angular clustering method based on image processing is proposed in this paper. It is used to identify clusters in 2D representations of propagation channels. The approach uses operations such as watershed segmentation and is particularly well suited for clustering directional channels obtained by beam-steering at millimeter-wave. This situation occurs for instance with electronic beam-steering using analog antenna arrays during beam training process or during channel modeling measurements using either electronic or mechanical beam-steering. In particular, the proposed technique is used here to cluster two-dimensional power angular spectrum maps. The proposed clustering is unsupervised and is well suited to preserve the shape of clusters by considering the angular connection between neighbor samples, which is useful to obtain more accurate descriptions of channel angular properties. The approach is found to outperform approaches based on K-Power- Means in terms of accuracy as well as computational resource . The technique is assessed in simulation using IEEE 802.11ad channel model and in measurement using experiments conducted at 60 GHz in an indoor environment.

eess.SP↗

arXiv2019

A detailed study on the intrinsic time resolution of the future MRPC detector

A challenging goal of 20 ps for the total time resolution of the Multi-gap Resistive Plate Chambers (MRPCs) system has been recently addressed by high energy physics experiments. In order to meet this requirement, one should have a deeper understanding of the detector physics and the factors that contribute to the intrinsic time resolution of the MRPC, which are studied in this paper. The sources of the timing uncertainties, their influences and the relationship with the detector geometry are analyzed. A comparison between the time resolutions obtained using two different simulation algorithms is presented. The obtained results are useful for the design of the timing MRPCs with an improved performance, and meanwhile offer guidance for the R&D activities in the field.

physics.ins-det↗

arXiv2018

The study of a new time reconstruction method for MRPC read out by waveform digitizer

The measurement of the $K^{\pm}$ production in the Semi-Inclusive Deep Inelastic Scattering (SIDIS) can provide further knowledge about the structure of nucleon, and thus it is purposed in the Solenoidal Large Intensity Device(SoLID) at Jefferson Lab(JLab). In this experiment, the identification of the kaons is planed to be accomplished with the Multi-gap Resistive Plate Chambers(MRPC), and the requirement for the time resolution is around 20 $\rm ps$. This is very challenging for the present MRPC systems (typical resolution 60 $\rm ps$), while in this paper, it is proved that the performance can be improved largely if the signal waveform is obtained and analyzed with a neural network method. In a cosmic ray experiment, the time resolution of a 6-gap 0.25$\rm mm$-thick MRPC reaches 36 $\rm ps$ with this method, and a even better performance is expected with a thinner MRPC.

physics.ins-det↗

arXiv2018

Status of technology of MRPC time of flight system

Time of flight system (TOF) based on MRPC technology is widely used in modern physics experiments, and it also plays an important role in particle identification. With the increase of accelerator energy and luminosity, TOF system is required to indentify definite particles precisely under high rate environment. The MRPC technology TOF system can be defined as three generations. The first generation TOF is based on float glass MRPC and its time resolution is around 80ps, but the rate is relatively low (typically lower than 100Hz/cm2). The typical systems are TOF of RHIC-STAR, LHC-ALICE and BES III endcap. For the second generation TOF, its time resolution is in the same order with the first generation, but the rate capability is much higher. Its rate capability can reach 30kHz/cm2. The typical experiment with this high rate TOF is FAIR-CBM. The biggest challenge is on the third generation TOF. For example, the momentum upper limit of K/PI separation is around 7GeV/c for JLab-SoLID TOF system under high particle rate as high as 20kHz/cm2, the time requirement is around 20ps. The readout electronics of first two generations is based on time over threshold method and pulse shape sampling technology will be used in the third generation TOF. In the same time, the machine learning technology is also designed to analysis the time performance. In this paper, we will describe the evolution of MRPC TOF technology and key technology of each generation TOF.

physics.ins-det↗

arXiv2016

Development and test of a real-size MRPC for CBM-TOF

In the CBM (Compressed Baryonic Matter) experiment constructed at the Facility for Anti-proton and Ion Research (Fair) at GSI, Darmstadt, Germany, MRPC(Multi-gap Resistive Plate Chamber) is adopted to construct the large TOF (Time-of-Flight) system to achieve an unprecedented precision of hadron identification, benefiting from its good time resolution, relatively high efficiency and low building price. We have developed a kind of double-ended readout strip MRPC. It uses low resistive glass to keep good performance of time resolution under high-rate condition. The differential double stack structure of 2x4 gas gaps help to reduce the required high voltage to half. There are 24 strips on one counter, and each is 270mm long, 7mm wide and the interval is 3mm. Ground is placed onto the MRPC electrode and feed through is carefully designed to match the 100 Ohm impedance of PADI electronics. The prototype of this strip MRPC has been tested with cosmic ray, a 98% efficiency and 60ps time resolution is gotten. In order to further examine the performance of the detector working under higher particle flux rate, the prototype has been tested in the 2014 October GSI beam time and 2015 February CERN beam time. In both beam times a relatively high rate of 1 kHz/cm2 was obtained. The calibration is done with CBM ROOT. A couple of corrections has been considered in the calibration and analysis process (including time-walk correction, gain correction, strip alignment correction and velocity correction) to access actual counter performances such as efficiency and time resolution. An efficiency of 97% and time resolution of 48ps are obtained. All these results show that the real-size prototype is fully capable of the requirement of the CBM-TOF, and new designs such as self-sealing are modified into the strip counter prototype to obtain even better performance.

physics.ins-det↗