Low Model Accuracy Research Articles

A Robust Client Selection Mechanism for Federated Learning Environments

There is a exponential growth of data usage, specially due to the proliferation of connected applications with personalized models for different applications. In this context, Federated Learning (FL) emerges as a promising solution to enable collaborative model training while preserving the privacy and autonomy of participating clients. In a typical FL scenario, clients exhibit significant heterogeneity in terms of data distribution and hardware configurations. In this way, randomly sampling clients in each training round may not fully exploit the local updates from heterogeneous clients, resulting in lower model accuracy, slower convergence rate, degraded fairness, etc. In addition, malicious users could disseminate incorrect weights, which may decrease the accuracy of aggregated models and increase the time for convergence in FL. In this article, we introduce Resilience-aware Client Selection Mechanism for non-IID data and malicious clients in FL environment, called RICA. The proposed mechanism employs data size and entropy as criteria for client selection. In addition, RICA relies Centroid-Based Kernel Alignment (CKA) to identify and exclude potentially malicious clients. Our evaluation shows an improvement of 125% in Accuracy values in a scenario of malicious clients, which means the RICA+CKA demonstrates a more stable and resilient approach, reaching 90% accuracy in a few rounds compared to the default average approach, reached only around 30%. Therefore, results of the behavior of RICA+CKA in different datasets show the evaluation of different numbers of clients reaching around 90% while the other approach does not pass the 50% Accuracy.

Side-Scan Sonar Image Augmentation Method Based on CC-WGAN

The utilization of deep learning algorithms for side-scan sonar target detection is impeded by the restricted quantity and representativeness of side-scan sonar (SSS) samples. To address this issue, this paper proposes a method for image augmentation using a CC-WGAN network. First, the generator incorporates the Convolutional Block Attention Module (CBAM) to enhance the assimilation of global information and local features in the input images. This integration also improves stability and avoids mode collapse problems associated with the original Generative Adversarial Network. Subsequently, the CBAM is incorporated into the discriminator to facilitate a better understanding of the relevance and significance of input data, thereby enhancing the model’s generalization ability. Finally, based on this model, existing few-sample SSS images are augmented, and we utilize the augmented images for discrimination and detection with YOLOv5. The experimental results show that following training with the SSS dataset that is augmented by this network, the accuracy of target detection increased by 7.6%, validating the feasibility of our proposed method. This method presents a novel solution to the problem of low model accuracy in underwater target detection with side-scan sonar due to limited samples.

Method for Remaining Useful Life Prediction of Turbofan Engines Combining Adam Optimization-Based Self-Attention Mechanism with Temporal Convolutional Networks

Conducting the remaining useful life (RUL) prediction for an aircraft engines is of significant importance in enhancing aircraft operation safety and formulating reasonable maintenance plans. Addressing the issue of low prediction model accuracy due to traditional neural networks’ inability to fully extract key features, this paper proposes an engine RUL prediction model based on the adaptive moment estimation (Adam) optimized self-attention mechanism–temporal convolutional network (SAM-TCN) neural network. Firstly, the raw data monitored by sensors are normalized, and RUL labels are set. A sliding window is utilized for overlapping sampling of the data, capturing more temporal features while eliminating data dimensionality. Secondly, the SAM-TCN neural network prediction model is constructed. The temporal convolutional network (TCN) neural network is used to capture the temporal dependency between data, solving the mapping relationship of engine degradation characteristics. A self-attention mechanism (SAM) is employed to adaptively assign different weight contributions to different input features. In the experiments, the root mean square error (RMSE) values on four datasets are 11.50, 16.45, 11.62, and 15.47 respectively. These values indicate further reduction in errors compared to methods reported in other literature. Finally, the SAM-TCN prediction model is optimized using the Adam optimizer to improve the training effectiveness and convergence speed of the model. Experimental results demonstrate that the proposed method can effectively learn feature data, with prediction accuracy superior to other models.

Byzantine-robust Federated Learning via Cosine Similarity Aggregation

Federated Learning (FL) is proposed to train a machine learning model for clients with different training data. During the training of FL, a centralized server is usually employed to aggregate local models from clients iteratively. The aggregation process suffers from Byzantine attacks, where clients’ models could be maliciously modified by attackers to degrade the training performance. Existing defense aggregation solutions use distances or angles between different gradients to identify and eliminate malicious models from clients. However, they do not work well due to the high dimensional property of the machine learning model. Distance-based solutions cannot effectively identify attackers when the gradient direction of the model is maliciously tampered with. Angle-based solutions face the issue of low model accuracy for large models. In this paper, we propose Convolutional Kernel Angle-based Defense Aggregation (CKADA) to improve defense performance under various Byzantine attacks. The key of CKADA is to use the angle between convolutional kernels as the attack detection metric because the obtuse angle indicates the wrong training direction. CKADA calculates the angle between a client’s convolutional kernel gradients and the server’s convolutional kernel gradients as the attacker detection metric and eliminates convolutional kernel gradients of clients that create an obtuse angle to mitigate the impact of attackers on the model. We evaluate the performance of CKADA using AlexNet, ResNet-50, and GoogLeNet under two typical attacks. Simulation results show that CKADA mitigates the impact of Byzantine attacks and outperforms existing angle-based solutions and distance-based solutions by improving inference accuracy up to 67% and 89% respectively.

PHSI-RTDETR: A Lightweight Infrared Small Target Detection Algorithm Based on UAV Aerial Photography

To address the issues of low model accuracy caused by complex ground environments and uneven target scales and high computational complexity in unmanned aerial vehicle (UAV) aerial infrared image target detection, this study proposes a lightweight UAV aerial infrared small target detection algorithm called PHSI-RTDETR. Initially, an improved backbone feature extraction network is designed using the lightweight RPConv-Block module proposed in this paper, which effectively captures small target features, significantly reducing the model complexity and computational burden while improving accuracy. Subsequently, the HiLo attention mechanism is combined with an intra-scale feature interaction module to form an AIFI-HiLo module, which is integrated into a hybrid encoder to enhance the focus of the model on dense targets, reducing the rates of missed and false detections. Moreover, the slimneck-SSFF architecture is introduced as the cross-scale feature fusion architecture of the model, utilizing GSConv and VoVGSCSP modules to enhance adaptability to infrared targets of various scales, producing more semantic information while reducing network computations. Finally, the original GIoU loss is replaced with the Inner-GIoU loss, which uses a scaling factor to control auxiliary bounding boxes to speed up convergence and improve detection accuracy for small targets. The experimental results show that, compared to RT-DETR, PHSI-RTDETR reduces model parameters by 30.55% and floating-point operations by 17.10%. Moreover, detection precision and speed are increased by 3.81% and 13.39%, respectively, and mAP50, impressively, reaches 82.58%, demonstrating the great potential of this model for drone infrared small target detection.

Communication-efficient federated learning via personalized filter pruning

With the popularity of mobile devices and the continuous growth of interactive data, FL (Federated Learning) has gradually become an effective mean to address the problems of privacy leakage and data silos. However, due to the heterogeneity and imbalance of participants, FL faces many challenges, including model accuracy, security, heterogeneous devices and data, privacy preservation, as well as communication overhead and efficiency. To address challenges such as high communication overhead and low model accuracy in FL, we innovatively introduce model pruning into the FL framework and propose a personalized filter pruning-based FL method named PF2 Learning (Personalized Filter Pruning Federal Learning). This method achieves reasonable filter pruning for all local models by performing personalized pruning on each local model. Specifically, on the device side, we use a pruning strategy based on the norm geometric median, which also considers the order dependence between adjacent layers and evaluates the contribution of filters involved in pruning to optimize the pruning for local models at a unified pruning rate. On the server side, we calculate the unified pruning ratio of the model based on the contribution of participants' model filters and training errors to maintain the consistency of the participants' model structures. To validate the effectiveness of our proposed method, we conducted federated image classification tasks on ResNet-18, VGG-11, DenseNet-121, and InceptionNet-V1 models using CIFAR-10, FEMNIST, and ImageNet datasets. The experimental results show that PF2 Learning outperforms most FL pruning methods and exhibits better model performance and accuracy.

RefineDet with Improved Attention Block for Chest X-ray Image Lesion Detection

In recent years, artificial intelligence technology has been widely used to assist radiologists in diagnosing and analyzing medical images. The use of artificial intelligence technology can well assist doctors in the localization of lesions. However, the mainstream target detection models at this stage are difficult to be practically applied in medical systems because of factors such as the use of large backbone networks and high input resolution, which leads to low model accuracy and high consumption of computational resources. In this paper, we propose a fast detection speed and high accuracy of the lung lesion detection network IAB- RefineDet. By improving the channel and spatial attention mechanisms and introducing the improved attention module into RefineDet, the lesion detection accuracy is dramatically improved without significantly increasing the number of parameters. We conduct extensive experiments on VinDr-CXR, the world's largest publicly available chest radiograph detection dataset, and comparative experiments with existing mainstream target detection models. The experimental results show that IAB-RefineDet achieves a mAP of 16.23%, and the lesion detection performance is significantly better than mainstream deep learning models.

Depth linear discrimination-oriented feature selection method based on adaptive sine cosine algorithm for software defect prediction

Software Defect Prediction (SDP) plays a vital role in the software development life cycle as it helps identify and fix software defects. However, predicting software defects with irrelevant features and overlapping classes is challenging and can lead to lengthy training and low model accuracy. To address these challenges, this research introduces a novel Depth Linear Discrimination-Oriented Feature Selection Method based on Adaptive Sine Cosine Algorithm, named Depth Adaptive Sine Cosine Feature Selection (DASC-FS). DASC-FS integrates the Adaptive Sine Cosine Algorithm (ASCA) as a search algorithm to determine the relevant features and adopts Depth Linear Discriminant Analysis (D-LDA) to identify the discriminative features that maximize class separation. The paper proposes ASCA which is a metaheuristic algorithm meticulously designed to enhance the search capabilities of the standard Sine Cosine Algorithm (SCA). Combining the simplicity of the SCA with the efficiency of multiple mutation operators inspired by Genetic Algorithms (GA), ASCA enhances the diversity of the solutions and imparts remarkable adaptability to various situations. Furthermore, this study introduces a novel linear discriminant method, called Depth Linear Discriminant Analysis (D-LDA) to enhance the robustness of the original LDA. D-LDA systematically integrates the matrix depth concept into LDA, offering a systematic approach to address the challenges associated with scatter matrix estimation. As matrix depth measures how central or deep a particular matrix is within a distribution with respect to different directions, it is an efficient tool for computing a robust scatter matrix estimator that can handle outliers and complex data structures. The experimental results showed that DASC-FS consistently obtains the highest accuracy compared to most existing methods by integrating ASCA and D-LDA, thereby considering both accuracy optimization and class separation. The results also show that the use of multiple mutation operators in ASCA improves the search process capabilities. The results also show that the capacity of D-LDA to reduce data dimensionality and increase class separation yields highly competitive results compared to other LDAs. Finally, features related to code size and complexity have emerged as key factors for SDP because they consistently rank as important features across different classifiers and datasets. DASC-FS offers a valuable solution in domain knowledge for enhancing predictive accuracy and understanding factors contributing to software defects through enhanced search capabilities, robust scatter matrix estimation, and the ability to reduce data dimensionality.

Lightweight convolutional neural networks with context broadcast transformer for real-time semantic segmentation

With the increasing application of embedded mobile devices in various fields, lightweight real-time semantic segmentation systems have attracted more and more attention. Many current methods have successfully reduced the model's parameters, but they have led to low model accuracy, diminishing their practical value. In recent years, the Transformer architecture has achieved good results in many tasks, effectively capturing long-range dependencies and enhancing accuracy. However, the Transformer is not adept at extracting local features, and the model's computational cost is generally too high, hindering real-time inference implementation. We propose a lightweight semantic segmentation network called LCBFormer-Net, which embeds Transformer units between asymmetric encoders and decoders to fully leverage their advantages. On the encoder side, we design the Lightweight Multi-Fusion Unit (LMFU) and Partition Grouping Shuffle Channel Attention (PGSCA). The former fully utilizes input features, merging information multiple times through multiple branches and employing depthwise convolutions with dilation rate to further obtain sufficient features. The latter includes a lightweight grouped channel attention, better guide feature extraction. The Lightweight Context Broadcast Transformer (LCB Transformer) is the Transformer unit we designed, with a lightweight structure that significantly reduces GPU memory consumption. It also improves self-attention and feed-forward networks, enhancing the model's robustness. The decoder includes the Multi-scale Semantic Information Attention Fusion (MSIAF) module, guiding the fusion of features at three different scales and employing a hybrid attention mechanism with both channel and spatial attention to guide feature extraction. LCBFormer-Net achieves good segmentation results with a parameter count of 0.88 M on multiple challenging datasets with diverse scenes.

Construction of a Sustainable Training System for Engineering and Technological Innovation Talents Based on CIPP Model in the Digital Era

With the advent of the digital era, the demand for engineering and technological talents in the market is increasing. It is crucial to design a sustainable training and evaluation system suitable for engineering and technological innovation talents. However, traditional data evaluation models have problems with unsuitable evaluation indicators and low model accuracy. Therefore, the paper selects the decision-making-oriented evaluation model to build the basic talent training evaluation indicators, and design expert questionnaires based on the Delphi method to screen indicators. Moreover, the paper also combines the adaptive genetic algorithm optimised backpropagation neural network to establish a talent cultivation evaluation model, and conducts simulation experiments using MATLAB to verify its feasibility. The results showed that the evaluation accuracy [Formula: see text] value of the research model was 0.99635, and the fitness value was 1.34, which was 0.5 higher than the unmodified model and achieved good evaluation results. At the same time, by comparing with the traditional genetic algorithm optimised model and the unimproved backpropagation model, the average evaluation accuracy of the research model increased by 66.44% and 13.59%, and the recall rate increased by 10.79% and 23.96%. The research model has improved the accuracy of evaluation, and its adaptability has also been enhanced, achieving superior evaluation results, which has important value in the cultivation of engineering and technological innovation talents.

Approach of establishing a high-resolution shading occupant behavior model in the office building

Occupant behavior in buildings is becoming a critical factor in determining indoor comfort and energy consumption performance. The diversity and uncertainty of occupant behavior affect the implementation of occupant-centric building operations. However, existing shading behavior models do not account for various shading states (fully open, fully closed, or partially open, significantly deviating from real-world scenarios), accessibility of shading behavior (ASB), room configurations (single-person rooms and multi-person rooms), occupancy, and other factors, resulting in lower model accuracy. This study aims to address these limitations by proposing a high-resolution shading occupant behavior (HRSOB) model that integrates occupancy considerations, accessibility of shading behavior, and data mining techniques. This HRSOB model can classify various shading occupant behaviors and be integrated into the optimization controls of occupant-centric buildings. The utilization of Markov chain occupancy models significantly improves the accuracy of predicting stochastic shading behavior. The ASB classification method analyzes the features of the occupant shading adjustment behavior (hereafter referred to as shading behavior) and objectively reflects shading behavior patterns using spatial features, thereby enhancing the resolution of the shading occupant behavior model. Data-driven techniques, such as feature selection and soft voting, enable the analysis, training, and application of shading behavior data, and rigorous evaluation and validation of the HRSOB model are conducted. The results demonstrated a 10.92% improvement in the predictive performance of the HRSOB model compared with the typical shading occupant behavior (SOB) model.

An industrial process fault diagnosis method based on independent slow feature analysis and stacked sparse autoencoder network

Deep learning, with its powerful multilayer nonlinear representation of deep neural networks, enables models trained based on deep learning to describe the true distribution of data more accurately and thus have better generalization capabilities. However, the training data are usually assumed to obey independent identical distribution. Industrial process data nowadays usually have complex characteristics such as non-Gaussian, nonlinear, dynamic, strong correlation, and high-dimensional, which make it difficult to meet the assumptions of deep learning. Using industrial process data directly would result in low model accuracy and bias in the model output. This paper proposes a fault diagnosis method based on independent slow feature analysis (ISFA) and stacked sparse autoencoder (SSAE) network. The statistical properties possessed by ISFA are used to perform preliminary feature extraction of industrial process data, and then the extracted preliminary features are input to the SSAE network for fine-grained feature extraction. The proposed two-step feature extraction method not only makes the data input to the SSAE network as close as possible to obey independent identical distribution and provides convincing fault diagnosis results, but also makes full use of the powerful feature extraction capability of the SSAE network, which makes the proposed method also significantly improve the detection accuracy of incipient faults. The feasibility and superiority of the proposed method is verified through the TE process.

Constructing convolutional neural network by utilizing nematode connectome: A brain-inspired method

Convolutional neural networks have achieved impressive results in areas such as computer vision tasks. Recently, more complex architectures have been designed that add additional operations and connections to the standard architecture to enable deeper network training. A large number of trainable parameters are often required to acquire accurate results. As the number of parameters increases, the network structure becomes more intricate and complex, resulting in low model accuracy and wasted computational resources. Motivated by the promising performance of brain-inspired neural computation principles, a new network structure can be constructed using a composition of the biological connectome. The nematode Caenorhabditis elegans (C. elegans), as the only organism whose connectome has been fully mapped, is suitable to be selected for the construction of neural network structures. In this paper, the neuron connectome of C. elegans is encapsulated as a connectome block and used as the connection pattern of the convolutional layer in the convolutional network. Therefore, this modified convolutional neural network structure is called the nematode connectome neural network (NCNN), and the detailed implementation is presented. Moreover, comparative experiments on the CIFAR-100, CIFAR-10, MNIST, and ImageNet datasets are conducted to fully demonstrate the effectiveness and feasibility of the proposed NCNN model on image classification. In addition, the success of the NCNN also provides new ideas for future network structure design. The source code is publicly available at https://github.com/LongJin-lab/Nematode-Connectome-Neural-Network.

Improved potato AGB estimates based on UAV RGB and hyperspectral images

Crops' above-ground biomass (AGB) is a crucial indicator that reflects crop health and predicts crop yield. However, using only optical vegetation indices (VIs) can produce inaccurate AGB estimates due to differences in crop varieties, growth stages, and measurement environments. Given the advantages of unmanned aerial vehicle (UAV) RGB and hyperspectral image fusion, this study evaluated the performance of multi-source remote sensing data for estimating potato AGB at multiple growth stages. In 2019, this study conducted potato trials with different varieties, fertilization levels, and planting densities at the Xiaotangshan Experiment Base (Beijing). UAV image and AGB data of potato three main stages were obtained from ground survey work. High-frequency information of the potato canopy was extracted from RGB images using discrete wavelet transform (DWT). VIs and wavelet energy coefficients were extracted from hyperspectral images using continuous wavelet transform (CWT). The linear relationships between potato AGB with VIs, high-frequency information, and wavelet coefficients were analyzed. Potato AGB estimation models were constructed based on single and multiple types of variables using multiple stepwise regression (MSR) and random forest (RF) models, respectively. This work showed the following results: (i) High-frequency information and wavelet coefficients were more sensitive to potato multi-growth stage AGB than VIs, and the latter were the most sensitive. (ii) Using VIs, high-frequency information, or wavelet coefficients separately to estimate the potato multi-growth stage AGB resulted in higher error and lower model accuracy. (iii) Combining VIs with either high-frequency information or wavelet coefficients improved the accuracy of AGB estimation, which was further improved by combining high-frequency information with wavelet coefficients. (iv) Combining VIs with both high-frequency information and wavelet coefficients provided the highest estimation accuracy using the MSR method. This combined AGB estimation model reduced the RMSE by 27%, 21%, and 16%, respectively, relative to VIs, high-frequency information, or wavelet coefficients alone. This result shows that the complementary advantages of multi-source UAV data can solve the challenge of insufficient AGB estimation by optical remote sensing. The work in this study provides remote sensing technology support to achieve potato crop growth monitoring and improve yield predictions.

An Improved Data-Driven Method for Steering Feedback Torque of Driving Simulator

Traditional modeling methods of steering feedback torque (SFT) widely used in driving simulator are mainly physics-based, but suffers from high model complexity and low model accuracy due to the inaccurate or unknown structure, dynamics, and related parameters of the steering system, such as backlash or dry friction. For these problems, data-driven methods were considered to have advantageous in modeling such a complex yet dynamic system. However, due to the limitation of the data collected for model training from real vehicle test, the commonly used artificial neural network (ANN) cannot fully reflect the steering dynamics, and does not work well in many cases where collected data do not cover, and thus lead to poor generalization performance. In this article, a long short-term memory (LSTM) model is proposed and proved to simulate well on the dynamics of SFT, and works well in most driving conditions. A driving simulator is developed to verify the performance of different models in terms of generalization of working conditions. Based on the driving simulator data, we found that ANN failure generally arises in the transition between conventional and unconventional operating conditions, which is recognized by the proposed driving pattern recognition method. The LSTM-based modeling method can effectively solve this problem. The comparative analysis reveals that the improved LSTM model can simulate the SFT with higher accuracy and broader adaptability to driving simulator conditions.

Hierarchical Automated Machine Learning Approach for Self-Optimizable Driving Distraction Recognition Based on Driver’s Lane-Keeping Performance

With the enrichment of smartphone uses, phone-related driving distractions have become a threat to driving safety. One way to mitigate driving distractions is to detect them and provide real-time warnings. However, most existing driving distraction recognition algorithms are pretrained models composed of structures, hyperparameters, and parameters that may not be able to account for drivers’ individual differences and, thus, might result in low model accuracy. This study proposes a domain-specific hierarchical automated machine learning (HAT-ML) model that self-learns personalized optimal models to detect driving distractions from vehicle movement data. The HAT-ML model integrates key modeling steps into auto-optimizable layers, including knowledge-based feature extraction, feature selection by recursive feature elimination, automated algorithm selection, and hyperparameter autotuning by Bayesian optimization. In our eight-degrees-of-freedom driving simulator experiment, we demonstrated the effectiveness of the proposed model using three driving distraction tasks: browsing a short message, browsing a long message, and answering a phone call. The HAT-ML model was found to be reliable and robust for predicting phone-related driving distraction, achieving satisfactory results with a predictive accuracy of 80% at the group level and 90% at the individual level. Moreover, the results revealed that each distraction and driver type required different optimized hyperparameter values, which demonstrated the value of utilizing HAT-ML to detect driving distractions. The key elements that dominated the performance of the model have several theoretical and practical implications. The proposed method not only enhanced performance, but also provided data-driven insights about model development.

Lightweight multi-scale attention-guided network for real-time semantic segmentation

The wide application of small mobile devices makes the demand for lightweight real-time semantic segmentation algorithm become more and more intense, which makes it become one of the most popular research topics in the field of computer vision. However, some current methods blindly pursue low parameter numbers and high inference speeds, resulting in excessively low model accuracy and a loss of practical value. Therefore, a lightweight multi-scale attention-guided network for real-time semantic segmentation(LMANet) based on asymmetric encoder-decoder is proposed in this paper to solve the above dilemmas. In the encoder, we propose multi-scale asymmetric residual(MAR) modules to extract local spatial information and context information to enhance feature expression. In the decoder, we design an attention feature fusion(AFF) module and an attention pyramid refining (APR) module. AFF module guides the fusion of low-level and middle-level feature information through high-level semantic information, and finally refines the fusion result through APR module. In addition, we improve the segmentation performance of the model with the help of the attention modules in the network. Our network is tested on two complex urban road datasets. The experimental results show that LMANet achieves 70.6% mIoU and 66.5% mIoU on Cityscapes and Camvid datasets at 112FPS and 333FPS respectively, only 0.95 M parameters without any pre-training or pre-processing. Compared with most of existing state-of-the-art models, our network not only guarantees reasonable inference speed and parameter quantity, but also improves the accuracy as much as possible, which makes it more practical.

A kinematic modeling scheme of three-axis "Satcom-on-the-Move" antenna based on modified Denavit-Hartenberg method

The current three-axes Satcom-on-the-Move(SOTM) antenna modeling method has some shortcomings, such as low model accuracy, poor portability and so on. In order to solve the above-mentioned shortcomings, a new modified Denavit-Hartenberg(NMDH) kinematics modeling scheme for the three-axes SOTM antenna was proposed. To overcome the modeling difficulties caused by the inherent mechanical structure of the antenna, and meet the requirements of the modified Denavit-Hartenberg(MDH) method, the virtual coordinate system and auxiliary coordinate systems are designed and added respectively on the basis of the MDH method, the forward kinematics model and inverse kinematics solution of the three-axes SOTM antenna are obtained. The correctness of the NMDH modeling scheme is verified by digital simulation. Finally, the system tests are carried out. The test results show that the NMDH modeling scheme proposed in this paper achieves good effect of antenna tracking satellite, and has stronger portability than the system identification modeling method commonly used in engineering.

A Decision Support Framework for Pollution Source Detection via Coupled Forward‐Inverse Optimization and Multi‐Information Fusion

AbstractDue to the uncertainty in sensor data, low model accuracy, and high parameter heterogeneity in water quality modeling, pollution source detection (PSD) typically results in a problem of multiple possible solutions, which is the so‐called non‐uniqueness effect. Identifying unique solution to PSD problems is fundamentally essential for water quality control in surface water and groundwater systems. This study proposes a decision support framework to reduce the impact of uncertainty and identify a unique solution using a consensus‐based multiple information fusion method. Multiple water quality information sources are fused in the framework via spatial clustering and temporal Bayesian updating. Considering the real‐world complexity, the recession‐curve displacement method is used to handle multi‐sources scenario to enhance the accuracy of the framework. Meanwhile, a coupled forward‐inverse model instead of random sampling is used to improve the solving efficiency of PSD. The framework is validated by a real water pollution event and a number of semi‐hypothetical case studies. The results show that the prediction accuracy of the single‐source and double‐source pollution problems are 88.40% and 82.69%, with an average relative error of 9.75% and 11.42%, respectively. In addition, the uncertainty contribution quantified by a variance decomposition approach suggests that the interaction between parameter and measurement uncertainties has the greatest impact on the PSD results. The results reveal the advantages of the proposed decision framework for government organizations and communities involved in water environment management.

PHY: A performance-driven hybrid communication compression method for distributed training

Distributed training is needed to shorten the training time of deep neural networks. However, the communication overhead often hurts performance efficiency, especially in a distributed computing environment with limited network bandwidth. Hence, gradient compression techniques have been proposed to reduce communication time. But, compression also has the risk of causing lower model accuracy and longer training time due to compression loss and compression time. As a result, compression may not consistently achieve desired results, and there are limited discussions on when and which compression should be used. To address this problem, we propose a performance-driven hybrid compression solution. We make three main contributions. (1) We describe a hybrid compression strategy that chooses the compression method for individual model gradients. (2) We build an offline performance estimator and an online loss monitor to ensure the compression decision can minimize training time without sacrificing mode accuracy. (3) Our implementation can be imported to existing deep learning frameworks and applicable to a wide range of compression methods. Up to 3.6x training performance speedup was observed compared to other state-of-the-art methods.