Search

In this article, the reference and probability matching priors for Cronbach's alpha will be derived. The performance of these two priors will be compared to that of the well-known Jeffreys prior and a divergence prior. Cronbach's alpha is a measure used to assess the reliability of a set of test items. A simulation study will be considered to compare the performance of the priors, where the coverage rates, average interval lengths, and standard deviations of the interval lengths will be computed. A second simulation study will be considered where the mean relative error will be compared for the various priors using three different loss functions. The following loss functions will be considered, Squared error loss, Absolute error loss, and Linex loss. An illustrative example will also be considered. Throughout the article, the random effects approach will be used. [ABSTRACT FROM AUTHOR]

Pansharpening is a traditional image fusion problem where the reference image (or ground truth) is not accessible. Machine-learning-based algorithms designed for this task require an extensive optimization phase of network parameters, which must be performed using unsupervised learning techniques. The learning phase can either rely on a companion problem where ground truth is available, such as by reproducing the task at a lower scale or using a pretext task, or it can use a reference-free cost function. This study focuses on the latter approach, where performance depends not only on the accuracy of the quality measure but also on the mathematical properties of these measures, which may introduce challenges related to computational complexity and optimization. The evaluation of the most recognized no-reference image quality measures led to the proposal of a novel criterion, the Regression-based QNR (RQNR), which has not been previously used. To mitigate computational challenges, an approximate version of the relevant indices was employed, simplifying the optimization of the cost functions. The effectiveness of the proposed cost functions was validated through the reduced-resolution assessment protocol applied to a public dataset (PairMax) containing images of diverse regions of the Earth's surface. [ABSTRACT FROM AUTHOR]

Memory has been the subject of scientific study for nearly 150 years. Because a broad range of studies have been done, we can now assess how effective memory is for a range of materials, from simple nonsense syllables to complex materials such as novels. Moreover, we can assess memory effectiveness for a variety of durations, anywhere from a few seconds up to decades later. Our aim here is to assess a range of factors that contribute to the patterns of retention and forgetting under various circumstances. This was done by taking a meta-analytic approach that assesses performance across a broad assortment of studies. Specifically, we assessed memory across 256 papers, involving 916 data sets (e.g., experiments and conditions). The results revealed that exponential-power, logarithmic, and linear functions best captured the widest range of data compared with power and hyperbolic-power functions. Given previous research on this topic, it was surprising that the power function was not the best-fitting function most often. Contrary to what would be expected, a substantial amount of data also revealed either stable memory over time or improvement. These findings can be used to improve our ability to model and predict the amount of information retained in memory. In addition, this analysis of a large set of memory data provides a foundation for expanding behavioral and neuroimaging research to better target areas of study that can inform the effectiveness of memory. [ABSTRACT FROM AUTHOR]

This paper presents a procedure for and evaluation of using a semantic similarity metric as a loss function for neural source code summarization. Code summarization is the task of writing natural language descriptions of source code. Neural code summarization refers to automated techniques for generating these descriptions using neural networks. Almost all current approaches involve neural networks as either standalone models or as part of a pretrained large language models, for example, GPT, Codex, and LLaMA. Yet almost all also use a categorical cross‐entropy (CCE) loss function for network optimization. Two problems with CCE are that (1) it computes loss over each word prediction one‐at‐a‐time, rather than evaluating a whole sentence, and (2) it requires a perfect prediction, leaving no room for partial credit for synonyms. In this paper, we extend our previous work on semantic similarity metrics to show a procedure for using semantic similarity as a loss function to alleviate this problem, and we evaluate this procedure in several settings in both metrics‐driven and human studies. In essence, we propose to use a semantic similarity metric to calculate loss over the whole output sentence prediction per training batch, rather than just loss for each word. We also propose to combine our loss with CCE for each word, which streamlines the training process compared to baselines. We evaluate our approach over several baselines and report improvement in the vast majority of conditions. [ABSTRACT FROM AUTHOR]

The detection of distribution box doors on construction sites is particularly important in site safety management, but the size and posture of distribution boxes vary in different scenarios, and there are still challenges. This article proposes an improved YOLOv8n construction site distribution box door status detection and recognition method. Firstly, Global Attention Mechanism is introduced to reduce information dispersion and enhance global interaction representation, preserving the correlation between spatial and channel information to strengthen the network's feature extraction capability during the detection process. Secondly, to tackle the problem of class imbalance in construction site distribution box door state detection, the Focal_EIoU detection box loss function is used to replace the CIoU loss function, optimizing the model's ability to learn from difficult samples.Lastly,the proposed method is evaluated on a dataset of distribution boxes with different shapes and sizes collected from various construction scenes. Experimental results demonstrate that the improved YOLOv8n algorithm achieves an average precision (mAP) of 82.1% at a speed of 66.7 frames per second, outperforming other classical object detection networks and the original network. This improved method provides an efficient and accurate solution for practical detection tasks in smart chemical sites, especially in enhancing feature extraction and processing difficult sample cases, which has made significant progress. [ABSTRACT FROM AUTHOR]

Loss functions are crucial in training generative adversarial networks (GANs) and shaping the resulting outputs. These functions, specifically designed for GANs, optimize generator and discriminator networks together but in opposite directions. GAN models, which typically handle large datasets, have been successful in the field of deep learning. However, exploring the factors that influence the success of GAN models developed for limited data problems is an important area of research. In this study, we conducted a comprehensive investigation into the loss functions commonly used in GAN literature, such as binary cross entropy (BCE), Wasserstein generative adversarial network (WGAN), least squares generative adversarial network (LSGAN), and hinge loss. Our research focused on examining the impact of these loss functions on improving output quality and ensuring training convergence in single-image GANs. Specifically, we evaluated the performance of a single-image GAN model, SinGAN, using these loss functions in terms of image quality and diversity. Our experimental results demonstrated that loss functions successfully produce high-quality, diverse images from a single training image. Additionally, we found that the WGAN-GP and LSGAN-GP loss functions are more effective for single-image GAN models.

Relative belief inferences are shown to arise as Bayes rules or limiting Bayes rules. These inferences are invariant under reparameterizations and possess a number of optimal properties. In particular, relative belief inferences are based on a direct measure of statistical evidence. [ABSTRACT FROM AUTHOR]

LR serves as a generalized linear classifier for binary classification in supervised learning, exhibiting characteristics of simplicity in structure, strong interpretability, and effective fitting when dealing with linear data. However, its classification performance becomes limited when confronted with high-dimensional, uncertain, and linearly inseparable data. To address the inherent limitations of logistic regression, this paper introduced the theory of granular computing and proposed a novel logistic regression model called rotating granular logistic regression (RGLR). This paper introduced the theory of rotating granulation, where different angles of rotation were applied to pairs of features forming a plane coordinate system. This process constructed rotating granules by rotating pairs of features at various angles on the plane coordinate system, and granulated to form rotating granule vectors on multiple plane coordinate systems. This paper further defined the size, measurement, and operational rules of granules, and proposed a loss function for rotating granular logistic regression. The optimized solution of the rotating granular logistic regression was obtained by solving the value of the loss function. Finally, experiments are conducted using multiple UCI datasets, and the results compared across various evaluation metrics, indicate the effectiveness of the rotating granular logistic regression model. [ABSTRACT FROM AUTHOR]

The potential of neuromorphic (NM) solutions often lies in their low-SWaP (Size, Weight, and Power) capabilities, which often drive their application to domains that could benefit from this. Nevertheless, spiking neural networks (SNNs), with their inherent time-based nature, present an attractive alternative also for areas where data features are present in the time dimension, such as time series forecasting. Time series data, characterized by seasonality and trends, can benefit from the unique processing capabilities of SNNs, which offer a novel approach for this type of task. Additionally, time series data can serve as a benchmark for evaluating SNN performance, providing a valuable alternative to traditional datasets. However, the challenge lies in the real-valued nature of time series data, which is not inherently suited for SNN processing. In this work, we propose a novel spike-encoding mechanism and two loss functions to address this challenge. Our encoding system, inspired by NM event-based sensors, converts the derivative of a signal into spikes, enhancing interoperability with the NM technology and also making the data suitable for SNN processing. Our loss functions then optimize the learning of subsequent spikes by the SNN. We train a simple SNN using SLAYER as a learning rule and conduct experiments using two electricity load forecasting datasets. Our results demonstrate that SNNs can effectively learn from encoded data, and our proposed DecodingLoss function consistently outperforms SLAYER's SpikeTime loss function. This underscores the potential of SNNs for time series forecasting and sets the stage for further research in this promising area of research. [ABSTRACT FROM AUTHOR]

Here, a machine-learning method based on a kinetically informed neural network (NN) is introduced. The proposed method is designed to analyze a time series of difference electron-density maps from a time-resolved X-ray crystallographic experiment. The method is named KINNTREX (kinetics-informed NN for time-resolved X-ray crystallography). To validate KINNTREX, multiple realistic scenarios were simulated with increasing levels of complexity. For the simulations, time-resolved X-ray data were generated that mimic data collected from the photocycle of the photoactive yellow protein. KINNTREX only requires the number of intermediates and approximate relaxation times (both obtained from a singular valued decomposition) and does not require an assumption of a candidate mechanism. It successfully predicts a consistent chemical kinetic mechanism, together with difference electron-density maps of the intermediates that appear during the reaction. These features make KINNTREX attractive for tackling a wide range of biomolecular questions. In addition, the versatility of KINNTREX can inspire more NN-based applications to time-resolved data from biological macromolecules obtained by other methods.

The construction of flood disaster loss functions is of great significance for understanding the development law of flood disaster loss and improving the timeliness of loss assessment. A method for constructing loss functions based on urban flood disaster events is proposed, which is coupled with dynamic simulation of spatial distribution of flood disaster factors, spatial processing of socio- economic data, calculation of direct and indirect disaster losses, and curve fitting. A case study was carried out in Shanghai to construct flood disaster loss- time functions, and the adjustment strategy of function parameters was established. The results show that under the rainfall scenario of Chicago rain type, the development process of direct economic losses from flood disasters shows an S-shaped curve and Weibull distribution function can be used as the loss function. The development process of daily loss from production reduction and shutdown shows an L-shaped curve and hyperbolic decline curve function can be used as the loss function. The development process of daily industry- related loss shows an inverted U- shaped curve and Hoerl distribution function can be used as the loss function. The cumulative industry-related loss is about 1. 77 times of the loss of production reduction or shutdown, and the indirect economic loss is about 1. 29 times of the direct economic loss. The parameters of the loss function and the probability of disaster occurrence exhibit a nonlinear monotone increasing or decreasing relationship, the relative error of flood loss assessment based on fitting parameters is less than 5% . The method in this study has been proven to have high simulation accuracy, providing a new way to understand the development law of flood disaster loss and improve the timeliness of flood disaster loss assessment. [ABSTRACT FROM AUTHOR]

In recent years, machine learning has become a frequently used method for statistical estimation. Random forest regression, decision tree regression, support vector regression and polynomial regression are commonly used supervised machine learning methods. The most commonly used loss function in gradient descent during the optimization phase of these methods is the quadratic loss function, which estimates model parameters by minimizing the cost. The selection of an appropriate loss function is crucial for method selection. There are several loss functions in the literature, such as absolute loss, logarithmic loss and squared error loss. In this study, we propose the use of an inverted normal loss function, which is a finite loss function, to gain a new perspective on minimizing cost and measuring performance in machine learning regression problems. We assert that this loss function provides more accurate estimations of cost minimisation as compared to the quadratic loss function, which is an infinite loss function. This article presents a new approach based on the inverted normal loss function for optimization in regression and performance metrics in machine learning. The procedure and its advantages are illustrated using a simulation study. [ABSTRACT FROM AUTHOR]

NLP or Natural Language Processing is a dramatically developing field. In the "new ordinary" forced by COVID-19, a huge extent of instructive material, news, and conversations occur through computerized media stages. This gives more text information accessible to work on! Initially, straightforward RNNS (Recurrent Neural Networks) were utilized for preparing text classification. Yet, as of late, there have been many new examination distributions that give better cutting-edge outcomes. One such is BERT. In this paper, I’ll put down my presentation of the BERT transformer utilizing different measurements. [ABSTRACT FROM AUTHOR]

Underwater images suffer from severe color attenuation and contrast reduction due to the poor and complex lighting conditions in the water. Most mainstream methods employing deep learning typically require extensive underwater paired training data, resulting in complex network structures, long training time, and high computational cost. To address this issue, a novel ZeroReference Parameter Estimation Network (Zero-UAE) model is proposed in this paper for the adaptive enhancement of underwater images. Based on the principle of light attenuation curves, an underwater adaptive curve model is designed to eliminate uneven underwater illumination and color bias. A lightweight parameter estimation network is designed to estimate dynamic parameters of underwater adaptive curve models. A tailored set of non-reference loss functions are developed for underwater scenarios to fine-tune underwater images, enhancing the network's generalization capabilities. These functions implicitly control the learning preferences of the network and effectively solve the problems of color bias and uneven illumination in underwater images without additional datasets. The proposed method examined on three widely used real-world underwater image enhancement datasets. Experimental results demonstrate that our method performs adaptive enhancement on underwater images. Meanwhile, the proposed method yields competitive performance compared with state-of-the-art other methods. Moreover, the Zero-UAE model requires only 17K parameters, minimizing the hardware requirements for underwater detection tasks. What'more, the adaptive enhancement capability of the Zero-UAE model offers a new solution for processing images under extreme underwater conditions, thus contributing to the advancement of underwater autonomous monitoring and ocean exploration technologies. [ABSTRACT FROM AUTHOR]

Heavily imbalanced datasets are common in lesion segmentation. Specifically, the lesions usually comprise less than 5% of the whole image volume when dealing with brain MRI. A common solution when training with a limited dataset is the use of specific loss functions that rebalance the effect of background and foreground voxels. These approaches are usually evaluated running a single cross-validation split without taking into account other possible random aspects that might affect the true improvement of the final metric (i.e., random weight initialisation or random shuffling). Furthermore, the evolution of the effect of the loss on the heavily imbalanced class is usually not analysed during the training phase. In this work, we present an analysis of different common loss metrics during training on public datasets dealing with brain lesion segmentation in heavy imbalanced datasets. In order to limit the effect of hyperparameter tuning and architecture, we chose a 3D Unet architecture due to its ability to provide good performance on different segmentation applications. We evaluated this framework on two public datasets and we observed that weighted losses have a similar performance on average, even though heavily weighting the gradient of the foreground class gives better performance in terms of true positive segmentation. [ABSTRACT FROM AUTHOR]

In this study, we present the Bayesian estimates of the unknown parameters of the Topp-Leone Lindley distribution using the maximum likelihood and Bayesian methods. In this study, the Bayes theorem was adopted for obtaining the posterior distribution of the shape parameter and scale parameter of the Topp-Leone Lindley distribution assuming the Jeffreys' (non-informative) prior for the shape parameter and the Gamma (conjugate) prior for the scale parameter under three different loss functions namely: Square Error Loss Function, Linear Exponential Loss Function and Generalized Entropy Loss Function. The posterior distribution derived for both parameters are not solvable analytically, it requires a numerical approximation techniques to obtain the solution. The Lindley approximation techniques was adopted to obtain the parameters of interest. The loss function were used to derive the estimates of both parameters with an assumption that the both parameters are unknown and independent. To ascertain the accuracy of these estimators, the proposed Bayesian estimators under different loss functions are compared with the corresponding maximum likelihood estimator using a Monte Carlo simulation on the performance of these estimators according to the mean square error and BIAS based on simulated samples simulated from the Topp-Leone Lindley distribution.. It was also observed for any fixed value of the parameters, as sample size increases, the mean square errors of the Bayesian Estimates and maximum likelihood estimates decrease. Also, the maximum likelihood estimates and Bayesian estimates converge to the same value as the sample gets larger except for Generalized Entropy Loss Function. [ABSTRACT FROM AUTHOR]

Bayesian nonparametric mixture models are widely used to cluster observations. However, one major drawback of the approach is that the estimated partition often presents unbalanced clusters' frequencies with only a few dominating clusters and a large number of sparsely-populated ones. This feature translates into results that are often uninterpretable unless we accept to ignore a relevant number of observations and clusters. Interpreting the posterior distribution as penalized likelihood, we show how the unbalance can be explained as a direct consequence of the cost functions involved in estimating the partition. In light of our findings, we propose a novel Bayesian estimator of the clustering configuration. The proposed estimator is equivalent to a post-processing procedure that reduces the number of sparsely-populated clusters and enhances interpretability. The procedure takes the form of entropy-regularization of the Bayesian estimate. While being computationally convenient with respect to alternative strategies, it is also theoretically justified as a correction to the Bayesian loss function used for point estimation and, as such, can be applied to any posterior distribution of clusters, regardless of the specific model used. [ABSTRACT FROM AUTHOR]

Reconstructing low-dose CT imaging deals with handling the inherent noise within the data, which makes it a complex mathematical problem known as an ill-posed inverse problem. Recent attention has shifted towards deep learning-based techniques in CT image reconstruction. However, these approaches encounter limitations due to extensive data requirements for training and validation. We propose an unsupervised CT reconstruction technique that leverages the power of Deep convolutional neural networks (Deep CNNs), demonstrating that a randomly initialized neural network can serve as a prior. We have proposed a completely unsupervised deep learning technique called Loss-construct unsupervised network (LUNA) adjustment for low-dose CT image reconstruction. Our approach combines the traditional reconstruction technique, i.e., simultaneous algebraic reconstruction technique (SART), and integrates the image prior i.e., weighted total variation (WTV) regularization within the Deep CNN model. The overall reconstruction process is optimized using the alternating direction method of multipliers (ADMM) framework, that balances the neural network’s internal representation with the observed data, yielding a more consistent and accurate final image. The proposed method uses various loss functions to update the Deep CNN. The optimal update of the network depends on the various loss functions. Different loss functions, including perceptual loss, SSIM loss, WL2 loss, WTV loss, and sinogram loss, are used to guide the overall reconstruction. This approach effectively handles the constraints of data limitation of deep learning-based techniques, offering a robust and unsupervised solution for low-dose CT image reconstruction.

Anomaly detection is an important application in factory environments. The sounds emitted by a manufacturing machine during runtime can be indicative of either normal performance or of mechanical failure. Traditionally, cosine losses are frequently utilized in anomalous sound detection (ASD) algorithms. In this work, we evaluate cosine losses within the context and scope of Deep Metric Learning (DML), and we investigate various ways to modify and potentially boost their performance. The impact of each modification on the performance of ASD systems is studied under extensive experimental settings and ablations utilizing both the DCASE 2022 and DCASE 2023 machine ASD datasets. Additionally, under the framework of DML, we develop a key new insight into the inner workings of cosine losses, and we verify it with an ablation.

This research work introduces two novel loss functions, pattern-loss (POL) and label similarity-based instance modeling (LSIM), for improving the performance of multi-label classification using artificial neural network-based techniques. These loss functions incorporate additional optimization constraints based on the distribution of multi-label class patterns and the similarity of data instances. By integrating these patterns during the network training process, the trained model is tuned to align with the existing patterns in the training data. The proposed approach decomposes the loss function into two components: the cross entropy loss and the pattern loss derived from the distribution of class-label patterns. Experimental evaluations were conducted on eight standard datasets, comparing the proposed methods with three existing techniques.The results demonstrate the effectiveness of the proposed approach, with POL and LSIM consistently achieving superior accuracy performance compared to the benchmark methods.

Manually checking for defects on industrial parts such as steel surfaces is ineffective, error-prone, and can damage a company’s reputation. Current automated methods often lack accuracy or real-time detection capabilities. Early detection allows for timely corrective action, such as removing defective parts or adjusting production parameters which can streamline the manufacturing process and improve brand reputation, customer satisfaction, and compliance with industry standards. This paper presents MoL-YOLOv7 (MobileNet integrated with attention and loss function to You Only Look Once version 7), a deep learning model for accurate and real-time detection of steel defects. MoL-YOLOv7 modifies the YOLOv7 model by inserting a MobileNet block that reduces computational complexity while maintaining accuracy, allowing for faster detection. Adding the SimAm (Simple Parameter free Attention module) attention mechanism to the MobileNet block refines feature representations for complex tasks such as steel defect detection. Finally, replacing loss functions with EIoU (Effective IoU), WIoU (Wise-IoU), and SIoU (Scylla-IoU) improves the localization accuracy and addresses the class imbalance in the data. The modified model achieves high accuracy and real-time detection, enabling a streamlined defect detection process. Experimental results show that the modified model involving different loss functions used in this work achieves high accuracy, i.e. 0.5% to 3.5% higher than the original model YOLOv7. The superiority and validity of our modified model are demonstrated by comparison with other attention mechanisms and loss functions integrated into YOLOv7, and also on different texture datasets, putting forward a modified method to detect surface defects on steel strips in daily operations.

Classification applications in order to obtain the desired information from satellite images are one of the increasing areas of study. Remote sensing satellite images are very difficult to obtain, but nevertheless, they can be used in many different areas. In this context, satellite images have become an important data source for land use and land cover analysis. In this study, the public EuroSAT dataset, which is compatible with land cover and land use classification, has been used. There are a total of 27000 images in this dataset, with ten different class labels, such as industry, forest, and lake, and approximately 3000 images in each. In order to perform a more precise and efficient classification study, the images have first been subjected to data augmentation using various techniques and have been divided into a training, test, and validation set. In terms of classification, six different convolutional neural network (CNN) architectures and two different vision transformer architectures have been used. Three experiments have been carried out for the proposed system, and the selected models have been trained with three different loss functions, and the results have been evaluated. The first of the loss functions used is cross-entropy, which is one of the most well-known and basic functions; the second is label smoothing cross-entropy; and the third is a new custom loss function that has been improved and consists of a special combination of both loss functions. The eight proposed models have been trained for all three loss functions; performance evaluations have been made by analyzing metrics such as accuracy, precision, and recall; and comparisons have been realized. After the proposed models have been evaluated in terms of these metrics, a Fuzzy Cognitive Map (FCM) has been created to integrate the model outputs to obtain the final classification result. For this construction, the accuracy of each model has been used in the neighborhood matrix of FCM. The accuracy of the proposed FCM method has been calculated as 99.97%. When the results of the proposed different models have been examined, it has been clearly seen that the new loss function improved in this study has been more successful than the other two loss functions in all models. In addition, the simulation results of FCM-based integration indicate that our approach has very high accuracy.

Enhancing the crop yield predictability in intercropping systems is important for optimizing agricultural productivity. However, accurately predicting yield in such systems is quite challenging due to complex interactions between crops. This study introduces an advanced methodology using integrated loss functions within an optimally driven Feedback Neural Network (FNN) approach to improve yield prediction in a pea-cucumber intercropping systems. Traditional models relying only on Mean Square Error (MSE) loss function often unable to capture the complexity of models, leading to suboptimal performance. To address this limitation, the advanced loss functions are introduced like Dynamic Margin Loss (DML), Risk-Adjusted Loss (RAL), Quantile Loss (QL), and Hybrid Agronomic Efficiency Loss (HAEL) along with three optimizers such as Adaptive Momentum (Adam), Root Mean Square Propagation (RMSprop), and Adaptive delta (Adadelta). These loss functions incorporate risk, uncertainty, and agronomic efficiency into the model training process, enhances predictive capabilities and robustness. This proposed framework is able to capture the complexity of yield prediction by incorporating agricultural factors. While Gradient Boost Machines (GBM) and Long Short Term Memory (LSTM) have some potential, they are not able to capture these dynamics. The sensitivity and weight analysis also focuses that HAEL targets important agronomic factors such as nitrogen uptake and residue biomass, which provide a holistic view of yield prediction. The proposed approach improves the predictive performance compared to traditional models and helps to identify the importance of features, which makes it an effective tool for decision making in sustainable agriculture. Selecting appropriate loss functions is essential to improve the accuracy and robustness of crops yield prediction models. Thus, study provides a strong foundation for enhancing yield prediction in intricate intercropping systems, which all significantly enhance the advancement of precision agriculture.

The potential of neuromorphic (NM) solutions often lies in their low-SWaP (Size, Weight, and Power) capabilities, which often drive their application to domains that could benefit from this. Nevertheless, spiking neural networks (SNNs), with their inherent time-based nature, present an attractive alternative also for areas where data features are present in the time dimension, such as time series forecasting. Time series data, characterized by seasonality and trends, can benefit from the unique processing capabilities of SNNs, which offer a novel approach for this type of task. Additionally, time series data can serve as a benchmark for evaluating SNN performance, providing a valuable alternative to traditional datasets. However, the challenge lies in the real-valued nature of time series data, which is not inherently suited for SNN processing. In this work, we propose a novel spike-encoding mechanism and two loss functions to address this challenge. Our encoding system, inspired by NM event-based sensors, converts the derivative of a signal into spikes, enhancing interoperability with the NM technology and also making the data suitable for SNN processing. Our loss functions then optimize the learning of subsequent spikes by the SNN. We train a simple SNN using SLAYER as a learning rule and conduct experiments using two electricity load forecasting datasets. Our results demonstrate that SNNs can effectively learn from encoded data, and our proposed DecodingLoss function consistently outperforms SLAYER’s SpikeTime loss function. This underscores the potential of SNNs for time series forecasting and sets the stage for further research in this promising area of research.

Incorporating higher-order optimization functions, such as Levenberg-Marquardt (LM) have revealed better generalizable solutions for deep learning problems. However, these higher-order optimization functions suffer from very large processing time and training complexity especially as training datasets become large, such as in multi-view classification problems, where finding global optima is a very costly problem. To solve this issue, we develop a solution for LM-enabled classification with, to the best of knowledge first-time implementation of hinge loss, for multiview classification. Hinge loss allows the neural network to converge faster and perform better than other loss functions such as logistic or square loss rates. We prove our method by experimenting with various multiclass classification challenges of varying complexity and training data size. The empirical results show the training time and accuracy rates achieved, highlighting how our method outperforms in all cases, especially when training time is limited. Our paper presents important results in the relationship between optimization and loss functions and how these can impact deep learning problems.

In recent years, deep learning methods have shown significant potential in soil moisture modeling. However, a prominent limitation of deep learning approaches has been the absence of physical mechanisms. To address this challenge, this study introduces two novel loss functions designed around physical mechanisms to guide deep learning models in capturing physical information within the data. These two loss functions are crafted to leverage the monotonic relationships between surface water variables and shallow soil moisture as well as deep soil water. Based on these physically-guided loss functions, two physically-guided Long Short-Term Memory (LSTM) networks, denoted as PHY-LSTM and PHYs-LSTM, are proposed. These networks are trained on the global ERA5-Land dataset, and the results indicate a notable performance improvement over traditional LSTM models. When used for global soil moisture forecasting for the upcoming day, PHY-LSTM and PHYs-LSTM models exhibit closely comparable results. In comparison to conventional data-driven LSTM models, both models display a substantial enhancement in various evaluation metrics. Specifically, PHYs-LSTM exhibits improvements in several key performance indicators: an increase of 13.6% in Kling-Gupta Efficiency (KGE), a 20.7% increase in Coefficient of Determination (R2), an 8.2% reduction in Root Mean Square Error (RMSE), and a 4.4% increase in correlation coefficient (R). PHY-LSTM also demonstrates improvements, with a 14.8% increase in KGE, a 19.6% increase in R2, an 8.2% reduction in RMSE, and a 4.4% increase in R. Additionally, both models exhibit enhanced physical consistency over a wide geographical area. Experimental results strongly emphasize that the incorporation of physical mechanisms can significantly bolster the predictive capabilities of data-driven soil moisture models.

This paper proposes a fabric defect detection algorithm based on the SA-Pix2pix network and transfer learning to address the issue of insufficient accuracy in detecting complex pattern fabric defects in scenarios with limited sample data. Its primary contribution lies in treating defects as disruptions to the fabric's texture. It leverages a generative adversarial network to reconstruct defective images, restoring them to images of normal fabric texture. Subsequently, the reconstituted images are subjected to dissimilarity calculations against defective images, leading to image segmentation for the purpose of defect detection. This approach addresses the issues of poor defect image reconstruction accuracy due to the limited ability of remote dependency modeling within the generator's convolutional neural network. It also tackles deficiencies in the generative adversarial network's loss function in handling image details. To enhance the structure and loss function of the generative adversarial network, it introduces self-attention mechanisms, L1 loss, and an improved structural loss, thus mitigating the problems of low defect image reconstruction accuracy and insufficient image detail handling by the network. To counteract the issue of declining model training accuracy in the face of sparse complex fabric defect samples, a channel-wise domain transfer learning approach is introduced. This approach constrains the training of the target network through feature distribution, thereby overcoming the problem of target network overfitting caused by limited sample data. The study employs three methods to experimentally compare and investigate five distinct complex pattern fabric defects. The results demonstrate that, when compared to two other defect detection methods, the approach advocated in this paper exhibits superior detection accuracy in scenarios with limited sample data. [ABSTRACT FROM AUTHOR]

This work is devoted to improving the quality of segmentation of images of various scientific papers and legal acts by neural network models by training them using modified loss functions that take into account special features of images of the appropriate subject domain. The analysis of existing loss functions is carried out, and new functions are proposed that work both with the coordinates of bounding boxes and use information about the pixels of the input image. To assess the quality, a neural network segmentation model with modified loss functions is trained, and a theoretical assessment is carried out using a simulation experiment showing the convergence rate and segmentation error. As a result of the study, rapidly converging loss functions are created that improve the quality of document image segmentation using additional information about the input data. [ABSTRACT FROM AUTHOR]

Object detection remains a pivotal aspect of remote sensing image analysis, and recent strides in Earth observation technology coupled with convolutional neural networks (CNNs) have propelled the field forward. Despite advancements, challenges persist, especially in detecting objects across diverse scales and pinpointing small-sized targets. This paper introduces YOLO-SE, a novel YOLOv8-based network that innovatively addresses these challenges. First, the introduction of a lightweight convolution SEConv in lieu of standard convolutions reduces the network's parameter count, thereby expediting the detection process. To tackle multi-scale object detection, the paper proposes the SEF module, an enhancement based on SEConv. Second, an ingenious Efficient Multi-Scale Attention (EMA) mechanism is integrated into the network, forming the SPPFE module. This addition augments the network's feature extraction capabilities, adeptly handling challenges in multi-scale object detection. Furthermore, a dedicated prediction head for tiny object detection is incorporated, and the original detection head is replaced by a transformer prediction head. To address adverse gradients stemming from low-quality instances in the target detection training dataset, the paper introduces the Wise-IoU bounding box loss function. YOLO-SE showcases remarkable performance, achieving an average precision at IoU threshold 0.5 (AP50) of 86.5% on the optical remote sensing dataset SIMD. This represents a noteworthy 2.1% improvement over YOLOv8 and YOLO-SE outperforms the state-of-the-art model by 0.91%. In further validation, experiments on the NWPU VHR-10 dataset demonstrated YOLO-SE's superiority with an accuracy of 94.9%, surpassing that of YOLOv8 by 2.6%. The proposed advancements position YOLO-SE as a compelling solution in the realm of deep learning-based remote sensing image object detection. [ABSTRACT FROM AUTHOR]