Inference Model Research Articles

A sub-Earth-mass planet orbiting Barnard's star. No evidence of transits in TESS photometry

A sub-Earth-mass planet orbiting Barnard's star, designated as Barnard b, has recently been announced. At almost the same time, the first photometric data of Barnard's star by the Transit Exoplanet Survey Satellite (TESS) was released in Sector 80. We explore the possibility of emergent transits of Barnard b in TESS photometry. The detrended 2 min light curve appears to be flat, with a flux root mean square of 0.411 parts per thousand. Attempts of blind and informed transit curve model inference suggest no evidence of transiting Barnard b, or any other body. This provides a 3sigma upper bound of 87.9 degrees for the orbital inclination of Barnard b.

Cervical Cancer 2010-2019: An Upper Midwest Catchment of 40,000 Square Miles.

American Indian (AI) women have a higher incidence and mortality from cervical cancer than non-Hispanic White (NHW) women in the US. Our purpose is to detail the clinical events in the cervical cancer prevention continuum among the AI and White women with cervical cancer on the US frontier. A cancer center with a nearly 40,000 square-mile catchment area maintained a detailed cancer registry connected to the clinic records of all cervical cancer patients between 2010-2019. This catchment area provided records of both an AI and a White population. Descriptive and inferential statistics and modeling predictions detailed the prevention continuum. Among the 126 with cervical cancer, 20% were AI, and 78% were White. Sixty percent did not participate in cervical cancer screening within the 5 years before their diagnosis, and on average, 9.2 years passed since the last cervical cancer screening. 91% presented with symptoms, and most women presented with 2 or more symptoms. Thirteen percent underwent a colposcopic diagnostic step, significantly delaying the time to diagnosis compared to other diagnostic steps. Sixty-nine percent of the histopathologic diagnoses were squamous cell carcinoma, and 27% were adenocarcinoma. Forty-nine percent presented at stage I regardless of histopathology. Chemotherapy and radiation therapy were most commonly combined. Sixty-three percent of the population survived, and 42% survived at least 3 years from diagnosis. Younger age and earlier stages at diagnosis were the significant adjusted predictors of survival. Our detailed cervical cancer prevention continuum events provide new data questioning the use of colposcopy for women symptomatic at presentation.

Dynamic Recursive Logit Model for Vehicle Driving Route Choices and Path Inference With Incomplete Fixed Location Sensor Data

Dynamic Recursive Logit Model for Vehicle Driving Route Choices and Path Inference With Incomplete Fixed Location Sensor Data

Recurrent Multiscale Feature Modulation for Geometry Consistent Depth Learning.

The U-Net-like coarse-to-fine network design is currently the dominant choice for dense prediction tasks. Although this design can often achieve competitive performance, it suffers from some inherent limitations, such as training error propagation from low to high resolution and the dependency on the deeper and heavier backbones. To design an effective network that performs better, we instead propose Recurrent Multiscale Feature Modulation (R-MSFM), a new lightweight network design for self-supervised monocular depth estimation. R-MSFM extracts per-pixel features, builds a multiscale feature modulation module, and performs recurrent depth refinement through a parameter-shared decoder at a fixed resolution. This network design enables our R-MSFM to maintain a more lightweight architecture and fundamentally avoid error propagation caused by the coarse-to-fine design. Furthermore, we introduce the mask geometry consistency loss to facilitate our R-MSFM for geometry consistent depth learning. This loss penalizes the inconsistency of the estimated depths between adjacent views within the nonoccluded and nonstationary regions. Experimental results demonstrate the superiority of our proposed R-MSFM both at model size and inference speed, and show state-of-the-art results on two datasets: KITTI and Make3D.

Parallel implementation and performance of super-resolution generative adversarial network turbulence models for large-eddy simulation

Super-resolution (SR) generative adversarial networks (GANs) are promising for turbulence closure in large-eddy simulation (LES) due to their ability to accurately reconstruct high-resolution data from low-resolution fields. Current model training and inference strategies are not sufficiently mature for large-scale, distributed calculations due to the computational demands and often unstable training of SR-GANs, which limits the exploration of improved model structures, training strategies, and loss-function definitions. Integrating SR-GANs into LES solvers for inference-coupled simulations is also necessary to assess their a posteriori accuracy, stability, and cost. We investigate parallelization strategies for SR-GAN training and inference-coupled LES, focusing on computational performance and reconstruction accuracy. We examine distributed data-parallel training strategies for hybrid CPU–GPU node architectures and the associated influence of low-/high-resolution subbox size, global batch size, and discriminator accuracy. Accurate predictions require training subboxes that are sufficiently large relative to the Kolmogorov length scale. Care should be placed on the coupled effect of training batch size, learning rate, number of training subboxes, and discriminator’s learning capabilities. We introduce a data-parallel SR-GAN training and inference library for heterogeneous architectures that enables exchange between the LES solver and SR-GAN inference at runtime. We investigate the predictive accuracy and computational performance of this arrangement with particular focus on the overlap (halo) size required for accurate SR reconstruction. Similarly, a posteriori parallel scaling for efficient inference-coupled LES is constrained by the SR subdomain size, GPU utilization, and reconstruction accuracy. Based on these findings, we establish guidelines and best practices to optimize resource utilization and parallel acceleration of SR-GAN turbulence model training and inference-coupled LES calculations while maintaining predictive accuracy.

Assessment of time-series-derived no-flood references for sar-based Bayesian flood mapping

ABSTRACT The systematic mapping of flood events with Synthetic Aperture Radar (SAR) data is an area of growing importance. One global flood mapping algorithm utilized within the Copernicus Emergency Management Service is based upon a Bayesian Inference model that compares a SAR image to a simulated reference image representing no-flood conditions. This no-flood reference image is at present generated using a harmonic model trained using historic time series, thereby producing a backscatter image representing mean seasonal conditions. One known weakness of this approach is that it cannot account for changing environmental conditions from year to year, potentially causing an overestimation of flood extent during dry periods, snow and frost, or other effects causing lower-than normal backscatter. To minimize this detrimental effect, we introduce an exponential filter to estimate the no-flood reference image by weighting the most recent backscatter observations according to their time difference to the current SAR acquisition. We compare the performance of the new exponential filter model against the harmonic model using a novel time-series flood mapping assessment approach. First, we assess their predictions against the actual SAR image time series for the year 2023. Then, we analyze the false positive rate of the corresponding flood maps generated to ensure the robustness of the automated algorithm outside of flood events. Furthermore, we perform qualitative and quantitative analyses of flood maps matching with semi-automatic results from Copernicus Emergency Management Services and Sentinel Asia as a reference. Our time-series analysis confirms increased false positive rates due to well-known environmental drivers and highlights issues with agricultural overestimation. In this regard, the time-series comparisons of the no-flood reference models show a clear improvement in the TU Wien algorithm with the exponential filter, effectively reducing false positive rates on non-flooded scenes in most study sites. The exponential filter performed better than the harmonic model in most flooded scenes, where sites show generally improved Critical Success Index and User’s accuracy. However, the exponential filter model has difficulties with sites with prolonged floods in the time series, requiring further development. Overall, the exponential filter no-flood reference model shows great promise for improved global near-real-time flood mapping.

PKDGAN: Private Knowledge Distillation With Generative Adversarial Networks

The deployment of deep learning applications has to address the increasing privacy concerns when using private and sensitive data for training. A conventional deep learning model is prone to privacy attacks that can recover the sensitive information from either model parameters or accesses to the inference model. Recently, differential privacy (DP) has been proposed to offer provable privacy guarantees by randomizing the training process of neural networks. However, many approaches tend to provide the worst case privacy protection for model publishing, inevitably impairing the accuracy of the trained models. Thus, we present a novel private knowledge transfer strategy, where the private teacher trained on sensitive data is not publicly accessible but the student models can be released with privacy guarantees. In this paper, a three-player (teacher-student-discriminator) learning framework, Private Knowledge Distillation with Generative Adversarial Networks (PKDGAN), is proposed, where the student acquires the distilled knowledge from the teacher and is trained with the discriminator to generate similar outputs as the teacher. Moreover, a cooperative learning strategy is also suggested to support the collective training of multiple students against the discriminator when each student is with insufficient unlabelled training data. To enforce rigorous privacy guarantees, PKDGAN applies a Rényi differential privacy mechanism throughout the training process, and use it with a moment accountant technique to track the privacy cost. PKDGAN allows students to be trained with unlabelled public data and very few epochs, which avoids the exposure of training data while ensuring model performance. In the experiments, PKDGAN is found to have consistently good performance on various datasets (MNIST, SVHN, CIFAR-10, and Market-1501). When compared to prior works [1], [2], PKDGAN exhibits 5-82% accuracy loss improvement without compromising any privacy guarantee.

Artificial Intelligence-Based Pathological Application to predict regional lymph node metastasis in Papillary Thyroid Cancer

In this study, a model for predicting lymph node metastasis in papillary thyroid cancer was trained using pathology images from the TCGA(The Cancer Genome Atlas) public dataset of papillary thyroid cancer, and a front-end inference model was trained using our center's dataset based on the concept of probabilistic propagation of nodes in graph neural networks. Effectively predicting whether a tumor will spread to regional lymph nodes using a single pathological image is the capacity of the model described above. This study demonstrates that regional lymph nodes in papillary thyroid cancer are a common and predictable occurrence, providing valuable ideas for future research. Now we publish the above research process and code for further study by other researchers, and we also make the above inference algorithm public at the URL: http:// thyroid-diseases-research.com/, with the hope that other researchers will validate it and provide us with ideas or datasets for further study.

Risk and burden of low birthweight related to maternal PM2.5 exposure in Iran: A national causal inference study

BackgroundMaternal ambient fine particulate matter (PM2.5) has been linked with fetal low birthweight (LBW), while national causal evidence is limited, particularly in low- and middle-income countries. MethodsWe used nationwide registry-based birth records of 3839531 singleton livebirths, with maternal age (MA) between 15 and 49 years, across 31 provinces in Iran from 2013 through 2018. Monthly gridded concentrations of ambient PM2.5 during entire pregnancy were estimated through validated spatiotemporal models. We adopted traditional logistic/linear regression and double robust model (DRM) based on generalized propensity scores (GPS) to evaluate associations of maternal PM2.5 exposure with LBW risk and birthweight change. MA-specific exposure-response functions were investigated using restricted cubic splines in both traditional and casual inference models. PM2.5-attributable LBW burden were parallelly estimated, under the scenarios of achieving counterfactual exposure levels of multiple WHO interim targets. ResultsA total of 228655 LBW births were identified, and mean birthweight of was 3187 g. Both traditional and causal modelling approaches yielded consistently positive associations of maternal PM2.5 exposure with increased LBW risk and reduced birthweight, but marginally larger effect estimates were seen under GPS-based DRM analyses. Specifically, a 10-μg/m3 increase in ambient PM2.5 during the entire pregnancy was associated with the odds of 1.071 (95% confidence interval: 1.066, 1.076) for LBW and reduced birthweight of 8.9 (8.4, 9.4) g in GPS-based DRM analysis. MA-stratified analyses suggested non-linear exposure-response relationships of PM2.5 exposure with both LBW (Pnonlinear <0.001) risk and birthweight change (Pnonlinear <0.001) over the exposure range of 17–104 μg/m3, with notably steeper slopes at low concentrations (<37.3 μg/m3). According to GPS-based DRM analysis, achieving the interim target level 1–3 of annual PM2.5 (i.e., 35, 25, and 15 μg/m3) could avoid 5.5%–32.6% of overall LBW infants and birthweight loss of 7.0–57.6 g per livebirth during 2013–2018 in Iran. ConclusionsOur findings provided nationwide evidence of causality for elevated LBW risk and birthweight loss associated with maternal PM2.5 exposure in Iran. Improving air quality through multisectoral efforts may avoid dramatic LBW burden to promote neonatal health later in life.

Automatic real-time crack detection using lightweight deep learning models

Crack detection methods using deep learning models such as convolutional neural network (CNN) and the newly developed vision transformer (ViT) are expanding. However, there is still a lack of comparative evaluation of these models in real-time crack detection. In this paper, a total of 14 lightweight deep learning models, comprising seven CNN models, five ViT models and two hybrid models, are trained to build deep learning-based crack detection methods. Comprehensive experiments are conducted on the publicly available DeepCrack dataset, including accuracy, inference time, robustness and transfer learning experiments to compare the effectiveness and real-time performance of models. In terms of accuracy metrics and robustness performance, the ViT model using SegFormer segmentation method with MiT-B1 as backbone has the best performance, and in terms of the model inference time, the ViT models using TopFormer segmentation method demonstrate the fastest performance. If both the accuracy and inference time are considered, TopFormer with its small version of the backbone network has relatively better real-time performance, while the ViT model using SegFormer segmentation method with MiT-B0 as backbone and the CNN model using the fully convolutional network (FCN) segmentation method with HRNetV2-W18-Small as backbone have higher mean intersection over union (mIoU) values on computers and mobile devices, respectively. We also find that pre-training on a dataset that is more relevant to the target application scenario rather than on the widely used ImageNet gives better results for deep learning models. This study provides a reference for engineers to make choices about lightweight deep learning models.

Dynamic Matrix Multiplication Using Reinforcement Learning

Efficient matrix multiplication is a core computational task in machine learning, deep learning, and scientific computing, as it significantly impacts the performance of model training and inference. This project explores the use of reinforcement learning (RL) to dynamically select the optimal matrix multiplication algorithm based on input matrix characteristics. We compare traditional matrix multiplication methods, including Strassen’s algorithm, Winograd’s algorithm, and tiled matrix multiplication, and implement a Q- learning-based agent to choose the most efficient algorithm adaptively. Our RL-driven approach enables real-time algorithm selection, balancing computational speed and resource usage across different matrix types and sizes. Experimental results demonstrate the system's ability to reduce execution time by selecting algorithms suited to specific scenarios, providing insights into algorithm efficiency and adaptability for various data sizes. Furthermore, this research suggests practical optimization strategies for handling large-scale computations in machine learning applications, including using parallel processing, GPU acceleration, and parameter tuning to minimize computational bottlenecks. This project underscores the potential of integrating RL for adaptive computation within AI workflows, offering a scalable solution for enhancing matrix operation efficiency in data-intensive applications

Implementation and Comparison of Coffee Bean Drying Temperature SettingsBased on Fuzzy Logic

This research examines the implementation and measurement of temperature in drying coffee beansusing a fuzzy logic approach. Two main methods, namely the Mamdani Method and the SugenoMethod, were applied and evaluated in this context. Data on temperature, humidity, and water contentof coffee beans are collected during the drying process for use in the implementation of both methods.Implementation is carried out using MATLAB software, with detailed steps for each method. TheMamdani method involves fuzzification processes, inference using fuzzy rules, and defuzzification toobtain concrete values for temperature settings. Meanwhile, the Sugeno Method also involvesfuzzification of input data, but uses a linear fuzzy model for inference, so it does not require adefuzzification stage. The results and discussion of this study highlight the performance differencesbetween the two methods. Evaluation is carried out based on temperature prediction accuracy andenergy efficiency. The Mamdani Method shows good accuracy in predicting temperature, while theSugeno Method highlights efficiency and efficiency in the temperature regulation process. Therefore,this study provides valuable insight into selecting a suitable method for temperature regulation ofcoffee bean dryers

Partially factorized variational inference for high-dimensional mixed models

Abstract While generalized linear mixed models are a fundamental tool in applied statistics, many specifications, such as those involving categorical factors with many levels or interaction terms, can be computationally challenging to estimate due to the need to compute or approximate high-dimensional integrals. Variational inference is a popular way to perform such computations, especially in the Bayesian context. However, naive use of such methods can provide unreliable uncertainty quantification. We show that this is indeed the case for mixed models, proving that standard mean-field variational inference dramatically underestimates posterior uncertainty in high-dimensions. We then show how appropriately relaxing the mean-field assumption leads to methods whose uncertainty quantification does not deteriorate in high-dimensions, and whose total computational cost scales linearly with the number of parameters and observations. Our theoretical and numerical results focus on mixed models with Gaussian or binomial likelihoods, and rely on connections to random graph theory to obtain sharp high-dimensional asymptotic analysis. We also provide generic results, which are of independent interest, relating the accuracy of variational inference to the convergence rate of the corresponding coordinate ascent algorithm that is used to find it. Our proposed methodology is implemented in the R package vglmer. Numerical results with simulated and real data examples illustrate the favourable computation cost versus accuracy trade-off of our approach compared to various alternatives.

THGB: predicting ligand-receptor interactions by combining tree boosting and histogram-based gradient boosting

Ligand-receptor interaction (LRI) prediction has great significance in biological and medical research and facilitates to infer and analyze cell-to-cell communication. However, wet experiments for new LRI discovery are costly and time-consuming. Here, we propose a computational model called THGB to uncover new LRIs. THGB first extracts feature information of Ligand-Receptor (LR) pairs using iFeature. Next, it adopts a tree boosting model to obtain representative LR features. Finally, it devises the histogram-based gradient boosting model to capture high-quality LRIs. To assess the THGB performance, we compared it with three new LRI prediction models (i.e., CellEnBoost, CellGiQ, and CellComNet) and one classical protein-protein interaction inference model PIPR. The results demonstrated that THGB achieved the best overall predictions in terms of six evaluation indictors (i.e., precision, recall, accuracy, F1-score, AUC, and AUPR). To measure the effect of LR feature selection on the prediction, THGB was compared with four feature selection methods (i.e., PCA, NMF, LLE, and TSVD). The results showed that the tree boosting model was more appropriate to select representative LR features and improve LRI prediction. We also conducted ablation study and found that THGB with feature selection outperformed THGB without feature selection. We hope that THGB is a useful tool to find new LRIs and further infer cell-to-cell communication.

Layered Query Retrieval: An Adaptive Framework for Retrieval-Augmented Generation in Complex Question Answering for Large Language Models

Retrieval-augmented generation (RAG) addresses the problem of knowledge cutoff and overcomes the inherent limitations of pre-trained language models by retrieving relevant information in real time. However, challenges related to efficiency and accuracy persist in current RAG strategies. A key issue is how to select appropriate methods for user queries of varying complexity dynamically. This study introduces a novel adaptive retrieval-augmented generation framework termed Layered Query Retrieval (LQR). The LQR framework focuses on query complexity classification, retrieval strategies, and relevance analysis, utilizing a custom-built training dataset to train smaller models that aid the large language model (LLM) in efficiently retrieving relevant information. A central technique in LQR is a semantic rule-based approach to distinguish between different levels of multi-hop queries. The process begins by parsing the user’s query for keywords, followed by a keyword-based document retrieval. Subsequently, we employ a natural language inference (NLI) model to assess whether the retrieved document is relevant to the query. We validated our approach on multiple single-hop and multi-hop datasets, demonstrating significant improvements in both accuracy and efficiency compared to existing single-step, multi-step, and adaptive methods. Our method exhibits high accuracy and efficiency, particularly on the HotpotQA dataset, where it outperforms the Adaptive-RAG method by improving accuracy by 9.4% and the F1 score by 16.14%. The proposed approach carefully balances retrieval efficiency with the accuracy of the LLM’s responses.

UUVDNet: An efficient unmanned underwater vehicle target detection network for multibeam forward-looking sonar

Precise localization of unmanned underwater vehicle (UUV) on multi-beam forward-looking sonar (MFLS) is a key technology in underwater robotic exploration. However, large appearance change and weak features of targets in MFLS image, pose a serious challenge on target detection. This paper proposes an efficient UUV detection model with an enhanced training scheme and convolutional block attention mechanism. The proposed training scheme is composed of dual-path gradient information optimization and adaptive loss adjustment, which improves the network’s detection capabilities and the underwater environment adaptability. The convolutional block attention module (CBAM) is developed for extracting channel and spatially weighted feature maps, and providing a mechanism to adaptive feature refinement. We construct a comprehensive sonar detection dataset from field experiments. By comparing existing state-of-the-art detection algorithms, our proposed UUVDNet outperforms Faster R-CNN, Cascade R-CNN, CenterNet, SSD, YOLOv8n, DPFIN, and MBSNN in terms of mAP50–95 by 39.5%, 35.7%, 8.4%, 14.0%, 2.5%, 13.6%, and 21.8%, respectively. Furthermore, our inference model size is the most compact among all detection models, with a size of 5.1M. Final ablation experiments effectively showcase the proposed components’ efficiency.

SIGRN: Inferring Gene Regulatory Network with Soft Introspective Variational Autoencoders

Gene regulatory networks (GRNs) exhibit the complex regulatory relationships among genes, which are essential for understanding developmental biology and uncovering the fundamental aspects of various biological phenomena. It is an effective and economical way to infer GRNs from single-cell RNA sequencing (scRNA-seq) with computational methods. Recent researches have been done on the problem by using variational autoencoder (VAE) and structural equation model (SEM). Due to the shortcoming of VAE generating poor-quality data, in this paper, a soft introspective adversarial gene regulatory network unsupervised inference model, called SIGRN, is proposed by introducing adversarial mechanism in building a variational autoencoder model. SIGRN applies “soft” introspective adversarial mode to avoid training additional neural networks and adding additional training parameters. It demonstrates superior inference accuracy across most benchmark datasets when compared to nine leading-edge methods. In addition, method SIGRN also achieves better performance on representing cells and generating scRNA-seq data in most datasets. All of which have been verified via substantial experiments. The SIGRN method shows promise for generating scRNA-seq data and inferring GRNs.

Bootstrap-based inference for multiple mean-variance changepoint models

Identifying multiple change points in the mean and/or variance is crucial across various fields, including finance and quality control. We introduce a novel technique that detects change points for the mean and/or variance of a noisy sequence and constructs confidence intervals for both the mean and variance of the sequence. This method integrates the weighted bootstrap with the Sequential Binary Segmentation (SBS) algorithm. Not only does our technique pinpoint the location and number of change points, but it also determines the type of change for each estimated point, specifying whether the change occurred in the mean, variance, or both. Our simulations show that our method outperforms other approaches in most scenarios, clearly demonstrating its superiority. Finally, we apply our technique to three datasets, including DNA copy number variation, stock volume, and traffic flow data, further validating its practical utility and wide-ranging applicability.

Explicating the Role of Agricultural Socialized Services on Chemical Fertilizer Use Reduction: Evidence from China Using a Double Machine Learning Model

Reducing chemical usage, particularly chemical fertilizers, is a crucial measure for advancing sustainable agricultural development. This study utilized field survey data from 894 maize farmers across three northeastern provinces of China. A double machine learning modeling framework was established to empirically examine the impact and mechanism of agricultural socialized services on chemical fertilizer use of farm households. The model addresses numerous stringent constraints of conventional causal inference models and effectively mitigates the “curse of dimensionality” issue. Current research indicates that agricultural socialized services can substantially decrease chemical fertilizer use among farmers. Further investigation reveals that these services facilitate this reduction by enhancing the mechanization level, promoting the use of organic fertilizers, and providing a labor substitution effect. The region heterogeneity test indicates that the impact of agricultural socialized services is more pronounced in Liaoning and Heilongjiang provinces geographically. Regarding the heterogeneity analysis of food crop income levels, agricultural socialized services can decrease chemical fertilizer use among farmers more effectively with elevated food crop income levels. Consequently, the findings imply that the socialization of agricultural services has substantial potential to be an effective chemical fertilizer reduction strategy to support the agricultural green transition, which can be enhanced through promoting the degree of mechanization, organic fertilizer application, and labor division and specialization.

Modeling Product Degradation with Heterogeneity: A General Random-Effects Wiener Process Approach

Degradation of many products in practical applications is often subject to unit-to-unit heterogeneity. Such heterogeneity can be attributed to heterogeneous quality of the raw materials and the fluctuation of the manufacturing process, as well as the heterogeneous usage conditions and environments. The heterogeneity leads to the scattering of the degradation rates and diffusion intensities in the population. To model this phenomenon, this study proposes a general random-effects Wiener process model that accounts for the unit-to-unit heterogeneity in the degradation drift and the volatility simultaneously. In particular, the drift of the Wiener process is characterized by a normal distribution and the diffusion parameter is characterized by an independent inverse Gaussian distribution. The proposed model is flexible for characterization of heterogeneous degradation, and permits an analytically tractable model inference. An EM algorithm incorporating the variational Bayesian method is developed to estimate the model parameters, and a parametric bootstrap approach is proposed to construct confidence intervals. The performance of the proposed model and the estimation algorithm is validated by Monte Carlo simulations. The degradation data of an infrared LED device and the wearing data of the magnetic head of a hard disk driver are studied based on the proposed model. With comprehensive comparative studies, the good performance of the proposed model in fitting the real degradation data is validated.