arXiv subjects

Lei Feng

Publications and source records attributed to Lei Feng.

At least 109 records · Page 6Linked to original sources

arXiv2025

Early Stopping Against Label Noise Without Validation Data

Early stopping methods in deep learning face the challenge of balancing the volume of training and validation data, especially in the presence of label noise. Concretely, sparing more data for validation from training data would limit the performance of the learned model, yet insufficient validation data could result in a sub-optimal selection of the desired model. In this paper, we propose a novel early stopping method called Label Wave, which does not require validation data for selecting the desired model in the presence of label noise. It works by tracking the changes in the model's predictions on the training set during the training process, aiming to halt training before the model unduly fits mislabeled data. This method is empirically supported by our observation that minimum fluctuations in predictions typically occur at the training epoch before the model excessively fits mislabeled data. Through extensive experiments, we show both the effectiveness of the Label Wave method across various settings and its capability to enhance the performance of existing methods for learning with noisy labels.

cs.LG↗

arXiv2025

Impact of Sub-MeV Dark Matter on the Cooling of Pulsating White Dwarfs

In our galaxy, white dwarfs inevitably undergo scattering and capture processes with the interstellar diffuse dark matter. The captured dark matter forms a dark halo that eventually evaporates or annihilates. Theoretical pulsation modes and observations of pulsating white dwarfs provide predictions about their evolution. This motivates us to study the impact of sub-MeV interstellar dark matter on the cooling processes of white dwarfs. In this work, we consider the collisions between dark matter and relativistic degenerate electrons inside white dwarfs, numerically calculating the energy input and output results from scattering, capture, evaporation, and annihilation processes. Based on observational data from G117-B15, we conclude that the maximum cooling luminosity of the interstellar sub-MeV dark matter is approximately $10^{22} \, \text{erg}/\text{s}$, which is insufficient to provide an effective cooling mechanism for white dwarfs. Finally, if future observations detect a pulsating white dwarf in the Galactic center, the potential sensitivity of this scenario could extend to the region$10^{-3}\,\text{MeV} < m_χ< 10 \, \text{MeV}$ and $6.02 \times 10^{-38}\,\text{cm}^2 > σ_0 \geq 1.5 \times 10^{40} \, \text{cm}^2$.

astro-ph.HE↗

arXiv2025

Cosmic ray-driven bioenergetics for Life in Molecular Clouds

According to models such as panspermia or the Nebula-Relay hypothesis, the ancestors of life on Earth once lived in molecular clouds. Then what are the energy source and bioenergetics for such lifeforms? In this paper, we propose a new bioenergetic mechanism powered by cosmic ray ionization of hydrogen molecules and we argue that it may relate to or be the origin of chemiosmosis. Based on this mechanism, we suggest that the Last Universal Common Ancestor is a type of lifeform that utilize hydrogen molecules as donors of electron transport chains.

astro-ph.EP↗

arXiv2025

Scheme to Detect the Strong-to-weak Symmetry Breaking via Randomized Measurements

Symmetry breaking plays a central role in classifying the phases of quantum many-body systems. Recent developments have highlighted a novel symmetry-breaking pattern, in which the strong symmetry of a density matrix spontaneously breaks to the week symmetry. This strong-to-weak symmetry breaking is typically detected using multi-replica correlation functions, such as the Rényi-2 correlator. In this letter, we propose a practical protocol for detecting strong-to-weak symmetry breaking in experiments using the randomized measurement toolbox. Our scheme involves collecting the results of random Pauli measurements for (i) the original quantum state and (ii) the quantum state after evolution with the charged operators. Based on the measurement results, with a large number of samples, we can obtain the exact solution to the Rényi-2 correlator. With a small sample size, we can still provide an alternative approach to estimate the phase boundary to a decent accuracy. We perform numerical simulations of Ising chains with all-to-all decoherence as an exemplary demonstration. Our result opens the opportunity for the experimental studies of the novel quantum phases in mixed quantum states.

quant-ph↗

arXiv2025

A possible evidence of pion condensation

This work demonstrates that once a large number of pion is condensed in a high-energy hadron collision, the gamma-ray spectrum from $π^0$ decay takes on a typical broken power-law shape, which has been documented in many astronomical observations, but we have not yet recognized it. We show that this pion condensation is caused by a large number of soft gluons condensed in protons.

hep-ph↗

arXiv2024

Does confidence calibration improve conformal prediction?

Conformal prediction is an emerging technique for uncertainty quantification that constructs prediction sets guaranteed to contain the true label with a predefined probability. Previous works often employ temperature scaling to calibrate classifiers, assuming that confidence calibration benefits conformal prediction. However, the specific impact of confidence calibration on conformal prediction remains underexplored. In this work, we make two key discoveries about the impact of confidence calibration methods on adaptive conformal prediction. Firstly, we empirically show that current confidence calibration methods (e.g., temperature scaling) typically lead to larger prediction sets in adaptive conformal prediction. Secondly, by investigating the role of temperature value, we observe that high-confidence predictions can enhance the efficiency of adaptive conformal prediction. Theoretically, we prove that predictions with higher confidence result in smaller prediction sets on expectation. This finding implies that the rescaling parameters in these calibration methods, when optimized with cross-entropy loss, might counteract the goal of generating efficient prediction sets. To address this issue, we propose Conformal Temperature Scaling (ConfTS), a variant of temperature scaling with a novel loss function designed to enhance the efficiency of prediction sets. This approach can be extended to optimize the parameters of other post-hoc methods of confidence calibration. Extensive experiments demonstrate that our method improves existing adaptive conformal prediction methods in classification tasks, especially with LLMs.

cs.LG↗

arXiv2024

JWST observations constrain the time evolution of fine structure constants and dark energy - electromagnetic coupling

It was hypothesized in the literature that some physical parameters may be time-evolving and the astrophysical data can serve as a probe. Recently, James Webb Space Telescope (JWST) have released its early observations. In this work, we select the JWST spectroscopic observations of the high redshift ($z>7.1$) galaxies with strong [OIII] ($λ=4959$ Å\,and $5007$ Å\,in the rest frame) emission lines to constraint the evolution of the fine structure constant ($α$). With the spectra from two galaxies at redshifts of $7.19$ and $8.47$, the deviation of $α$ to its fiducial value is found to be as small as $0.44^{+8.4+1.7}_{-8.3-1.7} \times 10^{-4}$ and $-10.0^{+18+1.5}_{-18-1.5} \times 10^{-4}$, respectively (the first error is statistical and the latter is systematic). The combination of our results with the previous data reveals that $\frac{1}α \frac{d α}{dt} = 0.30^{+4.5}_{-4.5} \times 10^{-17}~{\rm yr^{-1}}$. Clearly, there is no evidence for a cosmic evolution of $α$. The prospect of further constraining the time evolution of $α$ is also discussed. The scalar field of dark energy is hypothesized to drive the acceleration of the universe's expansion through an interaction with the electromagnetic field. By integrating the observational data of the fine-structure constant variation, $\frac{Δα}α(z)$, we have established a stringent upper limit on the coupling strength between dark energy and electromagnetism. Our analysis yields $ζ\leq 3.92 \times 10^{-7}$ at the 95\% confidence level, representing the most stringent bound to date.

astro-ph.CO↗

arXiv2024

Possibilities for methanogenic and acetogenic life in molecular cloud

According to panspermia, life on Earth may have originated from life forms transported through space from elsewhere. These life forms could have passed through molecular clouds, where the process of methanogenesis could have provided enough energy to sustain living organisms. In this study, we have calculated the Gibbs free energy released from synthesizing hydrocarbons for methanogenic (acetogenic) life in a molecular cloud, with methane (acetic acid) as the final metabolic product. Our calculations demonstrate that the chemical reactions during methanogenesis can release enough free energy to support living organisms. The methanogenic life may have served as the predecessor of life on Earth, and there is some preliminary evidence from various molecular biology studies to support this idea. Furthermore, we propose a potential distinguishing signal to test our model.

physics.pop-ph↗

arXiv2024

Realtime Particulate Matter and Bacteria Analysis of Peritoneal Dialysis Fluid using Digital Inline Holography

We developed a digital inline holography (DIH) system integrated with deep learning algorithms for real-time detection of particulate matter (PM) and bacterial contamination in peritoneal dialysis (PD) fluids. The system comprises a microfluidic sample delivery module and a DIH imaging module that captures holograms using a pulsed laser and a digital camera with a 40x objective. Our data processing pipeline enhances holograms, reconstructs images, and employs a YOLOv8n-based deep learning model for particle identification and classification, trained on labeled holograms of generic PD particles, Escherichia coli (E. coli), and Pseudomonas aeruginosa (P. aeruginosa). The system effectively detected and classified generic particles in sterile PD fluids, revealing diverse morphologies predominantly sized 1-5 um with an average concentration of 61 particles per microliter. In PD fluid samples spiked with high concentrations of E. coli and P. aeruginosa, our system achieved high sensitivity in detecting and classifying these bacteria at clinically relevant low false positive rates. Further validation against standard colony-forming unit (CFU) methods using PD fluid spiked with bacterial concentrations from approximately 100 to 10,000 bacteria per milliliter demonstrated a clear one-to-one correspondence between our measurements and CFU counts. Our DIH system provides a rapid, accurate alternative to traditional culture-based methods for assessing bacterial contamination in PD fluids. By enabling real-time sterility monitoring, it can significantly improve patient outcomes in PD treatment, facilitate point-of-care fluid production, reduce logistical challenges, and be extended to quality control in pharmaceutical production.

physics.optics↗

arXiv2024

Bayesian-guided Label Mapping for Visual Reprogramming

Visual reprogramming (VR) leverages the intrinsic capabilities of pretrained vision models by adapting their input or output interfaces to solve downstream tasks whose labels (i.e., downstream labels) might be totally different from the labels associated with the pretrained models (i.e., pretrained labels). When adapting the output interface, label mapping methods transform the pretrained labels to downstream labels by establishing a gradient-free one-to-one correspondence between the two sets of labels. However, in this paper, we reveal that one-to-one mappings may overlook the complex relationship between pretrained and downstream labels. Motivated by this observation, we propose a Bayesian-guided Label Mapping (BLM) method. BLM constructs an iteratively-updated probabilistic label mapping matrix, with each element quantifying a pairwise relationship between pretrained and downstream labels. The assignment of values to the constructed matrix is guided by Bayesian conditional probability, considering the joint distribution of the downstream labels and the labels predicted by the pretrained model on downstream samples. Experiments conducted on both pretrained vision models (e.g., ResNeXt) and vision-language models (e.g., CLIP) demonstrate the superior performance of BLM over existing label mapping methods. The success of BLM also offers a probabilistic lens through which to understand and analyze the effectiveness of VR. Our code is available at https://github.com/tmlr-group/BayesianLM.

cs.LG↗

arXiv2024

Generative AI Enabled Matching for 6G Multiple Access

In wireless networks, applying deep learning models to solve matching problems between different entities has become a mainstream and effective approach. However, the complex network topology in 6G multiple access presents significant challenges for the real-time performance and stability of matching generation. Generative artificial intelligence (GenAI) has demonstrated strong capabilities in graph feature extraction, exploration, and generation, offering potential for graph-structured matching generation. In this paper, we propose a GenAI-enabled matching generation framework to support 6G multiple access. Specifically, we first summarize the classical matching theory, discuss common GenAI models and applications from the perspective of matching generation. Then, we propose a framework based on generative diffusion models (GDMs) that iteratively denoises toward reward maximization to generate a matching strategy that meets specific requirements. Experimental results show that, compared to decision-based AI approaches, our framework can generate more effective matching strategies based on given conditions and predefined rewards, helping to solve complex problems in 6G multiple access, such as task allocation.

cs.NI↗

arXiv2024

Artificial intelligence inspired freeform optics design: a review

Integrating artificial intelligence (AI) techniques such as machine learning and deep learning into freeform optics design has significantly enhanced design efficiency, expanded the design space, and led to innovative solutions. This article reviews the latest developments in AI applications within this field, highlighting their roles in initial design generation, optimization, and performance prediction. It also addresses the benefits of AI, such as improved accuracy and performance, alongside challenges like data requirements, model interpretability, and computational complexity. Despite these challenges, the future of AI in freeform optics design looks promising, with potential advancements in hybrid design methods, interpretable AI, AI-driven manufacturing, and targeted research for specific applications. Collaboration among researchers, engineers, and designers is essential to fully harness AI's potential and drive innovation in optics.

cs.LG↗

arXiv2024

Prototype-based Optimal Transport for Out-of-Distribution Detection

Detecting Out-of-Distribution (OOD) inputs is crucial for improving the reliability of deep neural networks in the real-world deployment. In this paper, inspired by the inherent distribution shift between ID and OOD data, we propose a novel method that leverages optimal transport to measure the distribution discrepancy between test inputs and ID prototypes. The resulting transport costs are used to quantify the individual contribution of each test input to the overall discrepancy, serving as a desirable measure for OOD detection. To address the issue that solely relying on the transport costs to ID prototypes is inadequate for identifying OOD inputs closer to ID data, we generate virtual outliers to approximate the OOD region via linear extrapolation. By combining the transport costs to ID prototypes with the costs to virtual outliers, the detection of OOD data near ID data is emphasized, thereby enhancing the distinction between ID and OOD inputs. Experiments demonstrate the superiority of our method over state-of-the-art methods.

cs.CV↗

arXiv2024

The Spatial Distribution of $\rm CH_4$ and $\rm CO_2$ Ice around Protostars IRAS 16253-2429 and IRAS 23385+6053

The origin and evolution of organic molecules represent a pivotal issue in the fields of astrobiology and astrochemistry, potentially shedding light on the origins of life. The James Webb Space Telescope (JWST), with its exceptional sensitivity and spectral resolution, is well suitable to observe molecules such as methane ($\rm CH_4$). Our analysis focused on the distribution of $\rm CH_4$, $\rm CO_2$, $\rm H_2O$, $\rm{CH_3OH+NH_4^+}$ ice and silicate absorption dips at approximately 7.7, 15.0, 6.0, 6.7 and 10.0 micrometres in two protostars: IRAS 16253-2429 and IRAS 23385+6053. We extract the $\rm CH_4$, $\rm CO_2$, $\rm H_2O$, $\rm{CH_3OH+NH_4^+}$ ice equivalent width (EW) maps and silicate extinction maps of the two sources. Our results reveal that the spatial distribution of $\rm CH_4$ in the protostellar system IRAS 16253-2429 closely mirrors that of its $\rm CO_2$ ice, forming a surrounded distribution that encircles the central protostar. This alignment suggests a common formation mechanism and subsequent trapping within the protostellar envelope, which is consistent with the "Classical" dark-cloud chemistry with ion-molecule reaction. In contrast, the spatial distributions of various molecules in the system IRAS 23385+6053 exhibit low similarities, which may be attributed to the dynamic influences of outflows or accretion processes. These discrepancies highlight the complex interplay between physical processes and chemical evolution in protostellar environments.

astro-ph.GA↗

arXiv2024

Holographic Air-quality Monitor (HAM)

We introduce the holographic air-quality monitor (HAM) system, uniquely tailored for monitoring large particulate matter (PM) over 10 um in diameter, i.e., particles critical for disease transmission and public health but overlooked by most commercial PM sensors. The HAM system utilizes a lensless digital inline holography (DIH) sensor combined with a deep learning model, enabling real-time detection of PMs, with greater than 97% true positive rate at less than 0.6% false positive rate, and analysis of PMs by size and morphology at a sampling rate of 26 liters per minute (LPM), for a wide range of particle concentrations up to 4000 particles/L. Such throughput not only significantly outperforms traditional imaging-based sensors but also rivals some lower-fidelity, non-imaging sensors. Additionally, the HAM system is equipped with additional sensors for smaller PMs and various air quality conditions, ensuring a comprehensive assessment of indoor air quality. The performance of the DIH sensor within the HAM system was evaluated through comparison with brightfield microscopy, showing high concordance in size measurements. The efficacy of the DIH sensor was also demonstrated in two two-hour experiments under different environments simulating practical conditions with one involving distinct PM-generating events. These tests highlighted the HAM system's advanced capability to differentiate PM events from background noise and its exceptional sensitivity to irregular, large-sized PMs of low concentration.

physics.ins-det↗

arXiv2024

Exploring the Dark Energy Equation of State with JWST

Observations from the James Webb Space Telescope (JWST) have unveiled several galaxies with stellar masses $M_*\gtrsim10^{10} M_\odot$ at redshifts $7.4\lesssim z\lesssim 9.1$. These remarkable findings indicate an unexpectedly high stellar mass density, which contradicts the prediction of the $Λ\rm CDM$ model. Our study utilizes the Chevallier--Polarski--Linder (CPL) parameterization, one of the dynamic dark energy models, to probe the role of dark energy on shaping galaxy formation. By considering varying star formation efficiencies within this framework, our analysis demonstrates that in a universe with a higher proportion of dark energy, more massive galaxies are formed at high redshifts, given a fixed perturbation amplitude observed today. These intriguing results highlight the promising prospect of revealing the nature of dark energy by analyzing the high-redshift massive galaxies.

astro-ph.CO↗

arXiv2024

CommonUppRoad: A Framework of Formal Modelling, Verifying, Learning, and Visualisation of Autonomous Vehicles

Combining machine learning and formal methods (FMs) provides a possible solution to overcome the safety issue of autonomous driving (AD) vehicles. However, there are gaps to be bridged before this combination becomes practically applicable and useful. In an attempt to facilitate researchers in both FMs and AD areas, this paper proposes a framework that combines two well-known tools, namely CommonRoad and UPPAAL. On the one hand, CommonRoad can be enhanced by the rigorous semantics of models in UPPAAL, which enables a systematic and comprehensive understanding of the AD system's behaviour and thus strengthens the safety of the system. On the other hand, controllers synthesised by UPPAAL can be visualised by CommonRoad in real-world road networks, which facilitates AD vehicle designers greatly adopting formal models in system design. In this framework, we provide automatic model conversions between CommonRoad and UPPAAL. Therefore, users only need to program in Python and the framework takes care of the formal models, learning, and verification in the backend. We perform experiments to demonstrate the applicability of our framework in various AD scenarios, discuss the advantages of solving motion planning in our framework, and show the scalability limit and possible solutions.

cs.MA↗

arXiv2024

AsyCo: An Asymmetric Dual-task Co-training Model for Partial-label Learning

Partial-Label Learning (PLL) is a typical problem of weakly supervised learning, where each training instance is annotated with a set of candidate labels. Self-training PLL models achieve state-of-the-art performance but suffer from error accumulation problem caused by mistakenly disambiguated instances. Although co-training can alleviate this issue by training two networks simultaneously and allowing them to interact with each other, most existing co-training methods train two structurally identical networks with the same task, i.e., are symmetric, rendering it insufficient for them to correct each other due to their similar limitations. Therefore, in this paper, we propose an asymmetric dual-task co-training PLL model called AsyCo, which forces its two networks, i.e., a disambiguation network and an auxiliary network, to learn from different views explicitly by optimizing distinct tasks. Specifically, the disambiguation network is trained with self-training PLL task to learn label confidence, while the auxiliary network is trained in a supervised learning paradigm to learn from the noisy pairwise similarity labels that are constructed according to the learned label confidence. Finally, the error accumulation problem is mitigated via information distillation and confidence refinement. Extensive experiments on both uniform and instance-dependent partially labeled datasets demonstrate the effectiveness of AsyCo. The code is available at https://github.com/libeibeics/AsyCo.

cs.LG↗