Advances In Deep Learning Research Articles

Benchmarking and Categorizing the Performance of Neural Program Repair Systems for Java

Recent years have seen a rise in neural program repair systems in the software engineering community, which adopt advanced deep learning techniques to automatically fix bugs. Having a comprehensive understanding of existing systems can facilitate new improvements in this area and provide practical instructions for users. However, we observe two potential weaknesses in the current evaluation of NPR systems: ① published systems are trained with varying data, and ② NPR systems are roughly evaluated through the number of totally fixed bugs. Questions such as “what types of bugs are repairable for current systems” cannot be answered yet. Consequently, researchers can not make target improvements in this area and users have no idea of the real affair of existing systems. In this paper, we perform a systematic evaluation of the existing nine state-of-the-art NPR systems. To perform a fair and detailed comparison, we (1) build a new benchmark and framework that supports training and validating the nine systems with unified data, and (2) evaluate retrained systems with detailed performance analysis, especially on the effectiveness and the efficiency. We believe our benchmark tool and evaluation results could offer practitioners the real affairs of current NPR systems and the implications of further facilitating the improvements of NPR.

3D Character Animation and Asset Generation Using Deep Learning

Besides video content, a significant part of entertainment is represented by computer games and animations such as cartoons. Creating such entertainment is based on two fundamental steps: asset generation and character animation. The main problem stems from its repetitive nature and the needed amounts of concentration and skill. The latest advances in deep learning and generative techniques have provided a set of powerful tools which can be used to alleviate these problems by facilitating the tasks of artists and engineers and providing a better workflow. In this work we explore practical solutions for facilitating and hastening the creative process: character animation and asset generation. In character animation, the task is to either move the joints of a subject manually or to correct the noisy data coming out of motion capture. The main difficulties of these tasks are their repetitive nature and the needed amounts of concentration and skill. For the animation case, we propose two decoder-only transformer based solutions, inspired by the current success of GPT. The first, AnimGPT, targets the original animation workflow by predicting the next pose of an animation based on a set of previous poses, while the second, DenoiseAnimGPT, tackles the motion capture case by predicting the clean current pose based on all previous poses and the current noisy pose. Both models obtained good performances on the CMU motion dataset, with the generated results being imperceptible to the untrained human eye. Quantitative evaluation was performed using mean absolute error between the ground truth motion vectors and the predicted motion vector. For both networks AnimGPT and DenoiseAnimGPT errors were 0.345, respectively 0.2513 (for 50 frames) that indicates better performances compared with other solutions. For asset generation, diffusion models were used. Using image generation and outpainting, we created a method that generates good backgrounds by combining the idea of text conditioned generation and text conditioned image editing. A time coherent algorithm that creates animated effects for characters was obtained.

Multi-functional scar tissue discrimination platform construction and exploration of molecular mechanism for scar formation

Scars that form after skin injury can cause structural and functional skin damage. Currently, scar tissue determination relies mainly on doctors’ subjective observations and judgments and lacks objectivity. However, current deep learning models can only achieve specific discrimination using unimodal data, which limits the comprehensive understanding of scar tissue and may reduce accuracy and stability. To solve these problems, in this study, a skin scar recognition platform based on advanced deep learning and a weighted aggregation network fusion method is proposed. It is implemented using a residual network-based CNN model and a logistic regression model with L1 regularization and is suitable for both unimodal and multimodal data. The experimental results showed that the proposed platform achieved a satisfactory accuracy of 98.26% for image discrimination. In the gene discrimination model test performed on a test dataset containing 17 gene expression samples, all samples were accurately discriminated. In addition, the proposed multimodal discrimination model achieved a discrimination accuracy of 98.23%. These results validate the effectiveness of deep feature extraction and multimodal feature fusion techniques for image discrimination tasks. On this basis, to deeply explore the pathogenesis of scar formation, a method with the ability to integrate regularization, sparsity, and orthogonality constraints, multiconstraint joint non-negative matrix factorization (MCJNMF), was used to explore the genetic correlation between collagen micrographic image features and gene expression data. In this study, we confirmed the association between the calcium signaling pathway, MAPK signaling pathway, and collagen fiber repair, and successfully identified 11 potential therapeutic targets, including TRIM59 and TBC1D9, which provide important clues for future scar treatment and prevention strategies.

An End-to-End Deep Learning Framework for Fault Detection in Marine Machinery.

The Industrial Internet of Things has enabled the integration and analysis of vast volumes of data across various industries, with the maritime sector being no exception. Advances in cloud computing and deep learning (DL) are continuously reshaping the industry, particularly in optimizing maritime operations such as Predictive Maintenance (PdM). In this study, we propose a novel DL-based framework focusing on the fault detection task of PdM in marine operations, leveraging time-series data from sensors installed on shipboard machinery. The framework is designed as a scalable and cost-efficient software solution, encompassing all stages from data collection and pre-processing at the edge to the deployment and lifecycle management of DL models. The proposed DL architecture utilizes Graph Attention Networks (GATs) to extract spatio-temporal information from the time-series data and provides explainable predictions through a feature-wise scoring mechanism. Additionally, a custom evaluation metric with real-world applicability is employed, prioritizing both prediction accuracy and the timeliness of fault identification. To demonstrate the effectiveness of our framework, we conduct experiments on three types of open-source datasets relevant to PdM: electrical data, bearing datasets, and data from water circulation experiments.

Comprehensive Review on Natural Language Generation for Automated Report Writing in Finance

The financial industry is transforming with the advent of Natural Language Generation (NLG), a subset of Natural Language Processing (NLP), which automates data conversion into coherent and contextually relevant narratives. This paper presents a comprehensive review of NLG's application in financial report automation, tracing its evolution from template-based methods to advanced deep learning and knowledge graph techniques. We discuss the relevance of NLG in automating report generation, its role in enhancing data analysis and decision-making, and its potential to improve investor communications and compliance with regulations. The paper identifies research gaps, including the need for optimization, accuracy improvement, and the integration of machine learning models for better classification and prediction. A proposed methodology for structured report generation is outlined, leveraging deep learning architectures such as RNNs and LSTMs. Future work aims to address these gaps and further integrate NLG into financial reporting, promising to streamline processes, reduce costs, and provide more personalized and insightful financial narratives.

Exploring the Processing Paradigm of Input Data for End-to-End Deep Learning in Tool Condition Monitoring.

Tool condition monitoring technology is an indispensable part of intelligent manufacturing. Most current research focuses on complex signal processing techniques or advanced deep learning algorithms to improve prediction performance without fully leveraging the end-to-end advantages of deep learning. The challenge lies in transforming multi-sensor raw data into input data suitable for direct model feeding, all while minimizing data scale and preserving sufficient temporal interpretation of tool wear. However, there is no clear reference standard for this so far. In light of this, this paper innovatively explores the processing methods that transform raw data into input data for deep learning models, a process known as an input paradigm. This paper introduces three new input paradigms: the downsampling paradigm, the periodic paradigm, and the subsequence paradigm. Then an improved hybrid model that combines a convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) was employed to validate the model's performance. The subsequence paradigm demonstrated considerable superiority in prediction results based on the PHM2010 dataset, as the newly generated time series maintained the integrity of the raw data. Further investigation revealed that, with 120 subsequences and the temporal indicator being the maximum value, the model's mean absolute error (MAE) and root mean square error (RMSE) were the lowest after threefold cross-validation, outperforming several classical and contemporary methods. The methods explored in this paper provide references for designing input data for deep learning models, helping to enhance the end-to-end potential of deep learning models, and promoting the industrial deployment and practical application of tool condition monitoring systems.

Enhancing Melanoma Diagnosis with Advanced Deep Learning Models Focusing on Vision Transformer, Swin Transformer, and ConvNeXt.

Skin tumors, especially melanoma, which is highly aggressive and progresses quickly to other sites, are an issue in various parts of the world. Nevertheless, the one and only way to save lives is to detect it at its initial stages. This study explores the application of advanced deep learning models for classifying benign and malignant melanoma using dermoscopic images. The aim of the study is to enhance the accuracy and efficiency of melanoma diagnosis with the ConvNeXt, Vision Transformer (ViT) Base-16, and Swin Transformer V2 Small (Swin V2 S) deep learning models. The ConvNeXt model, which integrates principles of both convolutional neural networks and transformers, demonstrated superior performance, with balanced precision and recall metrics. The dataset, sourced from Kaggle, comprises 13,900 uniformly sized images, preprocessed to standardize the inputs for the models. Experimental results revealed that ConvNeXt achieved the highest diagnostic accuracy among the tested models. Experimental results revealed that ConvNeXt achieved an accuracy of 91.5%, with balanced precision and recall rates of 90.45% and 92.8% for benign cases, and 92.61% and 90.2% for malignant cases, respectively. The F1-scores for ConvNeXt were 91.61% for benign cases and 91.39% for malignant cases. This research points out the potential of hybrid deep learning architectures in medical image analysis, particularly for early melanoma detection.

Advancements in deep learning and natural language processing for effective disaster sentiment analysis: A review

This article comprehensively studies the application of deep learning (DL), natural language processing (NLP), and large language models (LLM) in sentiment analysis in disaster scenarios such as earthquakes and major accidents. The article focuses on the latest developments in these technologies and their role in strengthening disaster management and response. The article explores various methods, including BERT, LSTM, and convolutional neural networks, with a focus on their practicality, challenges, and potential for development. This review aims to provide researchers and relevant practitioners with a comprehensive understanding of this rapidly developing field.

RS-Xception: A Lightweight Network for Facial Expression Recognition

Facial expression recognition (FER) utilizes artificial intelligence for the detection and analysis of human faces, with significant applications across various scenarios. Our objective is to deploy the facial emotion recognition network on mobile devices and extend its application to diverse areas, including classroom effect monitoring, human–computer interaction, specialized training for athletes (such as in figure skating and rhythmic gymnastics), and actor emotion training. Recent studies have employed advanced deep learning models to address this task, though these models often encounter challenges like subpar performance and an excessive number of parameters that do not align with the requirements of FER for embedded devices. To tackle this issue, we have devised a lightweight network structure named RS-Xception, which is straightforward yet highly effective. Drawing on the strengths of ResNet and SENet, this network integrates elements from the Xception architecture. Our models have been trained on FER2013 datasets and demonstrate superior efficiency compared to conventional network models. Furthermore, we have assessed the model’s performance on the CK+, FER2013, and Bigfer2013 datasets, achieving accuracy rates of 97.13%, 69.02%, and 72.06%, respectively. Evaluation on the complex RAF-DB dataset yielded an accuracy rate of 82.98%. The incorporation of transfer learning notably enhanced the model’s accuracy, with a performance of 75.38% on the Bigfer2013 dataset, underscoring its significance in our research. In conclusion, our proposed model proves to be a viable solution for precise sentiment detection and estimation. In the future, our lightweight model may be deployed on embedded devices for research purposes.

Implementation of Large Language Models and Agricultural Knowledge Graphs for Efficient Plant Disease Detection

This study addresses the challenges of elaeagnus angustifolia disease detection in smart agriculture by developing a detection system that integrates advanced deep learning technologies, including Large Language Models (LLMs), Agricultural Knowledge Graphs (KGs), Graph Neural Networks (GNNs), representation learning, and neural-symbolic reasoning techniques. The system significantly enhances the accuracy and efficiency of disease detection through an innovative graph attention mechanism and optimized loss functions. Experimental results demonstrate that this system significantly outperforms traditional methods across key metrics such as precision, recall, and accuracy, with the graph attention mechanism excelling in all aspects, particularly achieving a precision of 0.94, a recall of 0.92, and an accuracy of 0.93. Furthermore, comparative experiments with various loss functions further validate the effectiveness of the graph attention loss mechanism in enhancing model performance. This research not only advances the application of deep learning in agricultural disease detection theoretically but also provides robust technological tools for disease management and decision support in actual agricultural production, showcasing broad application prospects and profound practical value.

A brief review of hypernetworks in deep learning

Hypernetworks, or hypernets for short, are neural networks that generate weights for another neural network, known as the target network. They have emerged as a powerful deep learning technique that allows for greater flexibility, adaptability, dynamism, faster training, information sharing, and model compression. Hypernets have shown promising results in a variety of deep learning problems, including continual learning, causal inference, transfer learning, weight pruning, uncertainty quantification, zero-shot learning, natural language processing, and reinforcement learning. Despite their success across different problem settings, there is currently no comprehensive review available to inform researchers about the latest developments and to assist in utilizing hypernets. To fill this gap, we review the progress in hypernets. We present an illustrative example of training deep neural networks using hypernets and propose categorizing hypernets based on five design criteria: inputs, outputs, variability of inputs and outputs, and the architecture of hypernets. We also review applications of hypernets across different deep learning problem settings, followed by a discussion of general scenarios where hypernets can be effectively employed. Finally, we discuss the challenges and future directions that remain underexplored in the field of hypernets. We believe that hypernetworks have the potential to revolutionize the field of deep learning. They offer a new way to design and train neural networks, and they have the potential to improve the performance of deep learning models on a variety of tasks. Through this review, we aim to inspire further advancements in deep learning through hypernetworks.

Label-free spatiotemporal decoding of single-cell fate via acoustic driven 3D tomography

Label-free three-dimensional imaging plays a crucial role in unraveling the complexities of cellular functions and interactions in biomedical research. Conventional single-cell optical tomography techniques offer affordability and the convenience of bypassing laborious cell labelling protocols. However, these methods are encumbered by restricted illumination scanning ranges on abaxial plane, resulting in the loss of intricate cellular imaging details. The ability to fully control cellular rotation across all angles has emerged as an optimal solution for capturing comprehensive structural details of cells. Here, we introduce a label-free, cost-effective, and readily fabricated contactless acoustic-induced vibration system, specifically designed to enable multi-degree-of-freedom rotation of cells, ultimately attaining stable in-situ rotation. Furthermore, by integrating this system with advanced deep learning technologies, we perform 3D reconstruction and morphological analysis on diverse cell types, thus validating groups of high-precision cell identification. Notably, long-term observation of cells reveals distinct features associated with drug-induced apoptosis in both cancerous and normal cells populations. This methodology, based on deep learning-enabled cell 3D reconstruction, charts a novel trajectory for groups of real-time cellular visualization, offering promising advancements in the realms of drug screening and post-single-cell analysis, thereby addressing potential clinical requisites.

Mitigating digital market risk with conventional, green, and Islamic bonds: Fresh insights from new hybrid deep learning models

We examine the impact of conventional, green, and Islamic bonds on the long-term memory of cryptocurrency market risk. Utilizing a time-varying parameter vector autoregressive deep learning model, we integrate time-varying parameter vector autoregressive methods with advanced deep learning sequence modeling architectures, including temporal convolutional network, gated recurrent unit, and long short-term memory for December 18, 2017, to April 19, 2024. Results indicate that incorporating all fixed-income securities reduces digital market risk. However, conventional and green bonds have a particularly strong impact on improving the long-term memory of digital market risk, while this is not the case for Sukuk.

Discovery of potential antidiabetic peptides using deep learning

Antidiabetic peptides (ADPs), peptides with potential antidiabetic activity, hold significant importance in the treatment and control of diabetes. Despite their therapeutic potential, the discovery and prediction of ADPs remain challenging due to limited data, the complex nature of peptide functions, and the expensive and time-consuming nature of traditional wet lab experiments. This study aims to address these challenges by exploring methods for the discovery and prediction of ADPs using advanced deep learning techniques. Specifically, we developed two models: a single-channel CNN and a three-channel neural network (CNN + RNN + Bi-LSTM). ADPs were primarily gathered from the BioDADPep database, alongside thousands of non-ADPs sourced from anticancer, antibacterial, and antiviral peptide datasets. Subsequently, data preprocessing was performed with the evolutionary scale model (ESM-2), followed by model training and evaluation through 10-fold cross-validation. Furthermore, this work collected a series of newly published ADPs as an independent test set through literature review, and found that the CNN model achieved the highest accuracy (90.48 %) in predicting the independent test set, surpassing existing ADP prediction tools. Finally, the application of the model was considered. SeqGAN was used to generate new candidate ADPs, followed by screening with the constructed CNN model. Selected peptides were then evaluated using physicochemical property prediction and structural forecasts for pharmaceutical potential. In summary, this study not only established robust ADP prediction models but also employed these models to screen a batch of potential ADPs, addressing a critical need in the field of peptide-based antidiabetic research.

Deep learning models for separate segmentations of intracerebral and intraventricular hemorrhage on head CT and segmentation quality assessment.

The volume measurement of intracerebral hemorrhage (ICH) and intraventricular hemorrhage (IVH) provides critical information for precise treatment of patients with spontaneous ICH but remains a big challenge, especially for IVH segmentation. However, the previously proposed ICH and IVH segmentation tools lack external validation and segmentation quality assessment. This study aimed to develop a robust deep learning model for the segmentation of ICH and IVH with external validation, and to provide quality assessment for IVH segmentation. In this study, a Residual Encoding Unet (REUnet) for the segmentation of ICH and IVH was developed using a dataset composed of 977 CT images (all contained ICH, and 338 contained IVH; a five-fold cross-validation procedure was adopted for training and internal validation), and externally tested using an independent dataset consisting of 375 CT images (all contained ICH, and 105 contained IVH). The performance of REUnet was compared with six other advanced deep learning models. Subsequently, three approaches, including Prototype Segmentation (ProtoSeg), Test Time Dropout (TTD), and Test Time Augmentation (TTA), were employed to derive segmentation quality scores in the absence of ground truth to provide a way to assess the segmentation quality in real practice. For ICH segmentation, the median (lower-quantile-upper quantile) of Dice scores obtained from REUnet were 0.932 (0.898-0.953) for internal validation and 0.888 (0.859-0.916) for external test, both of which were better than those of other models while comparable to that of nnUnet3D in external test. For IVH segmentation, the Dice scores obtained from REUnet were 0.826 (0.757-0.868) for internal validation and 0.777 (0.693-0.827) for external tests, which were better than those of all other models. The concordance correlation coefficients between the volumes estimated from the REUnet-generated segmentations and those from the manual segmentations for both ICH and IVH ranged from 0.944 to 0.987. For IVH segmentation quality assessment, the segmentation quality score derived from ProtoSeg was correlated with the Dice Score (Spearman r=0.752 for the external test) and performed better than those from TTD (Spearman r=0.718) and TTA (Spearman r=0.260) in the external test. By setting a threshold to the segmentation quality score, we were able to identify low-quality IVH segmentation results by ProtoSeg. The proposed REUnet offers a promising tool for accurate and automated segmentation of ICH and IVH, and for effective IVH segmentation quality assessment, and thus exhibits the potential to facilitate therapeutic decision-making for patients with spontaneous ICH in clinical practice.

Monitoring activity index and behaviors of cage-free hens with advanced deep learning technologies

Chickens’ behaviors and activities are important information for managing animal health and welfare in commercial poultry houses. In this study, convolutional neural networks (CNNs) models were developed to monitor the chicken activity index. A dataset consisting of 1,500 top-view images was utilized to construct tracking models, with 900 images allocated for training, 300 for validation, and 300 for testing. Six different CNN models were developed, based on YOLOv5, YOLOv8, ByteTrack, DeepSORT, and StrongSORT. The final results demonstrated that the combination of YOLOv8 and DeepSORT exhibited the highest performance, achieving a Multi-Object Tracking Accuracy (MOTA) of 94%. Further application of the optimal model could facilitate the detection of abnormal behaviors such as smothering and piling, and enabled the quantification of flock activity into three levels (low, medium, and high) to evaluate footpad health states in the flock. This research underscores the application of deep learning in monitoring poultry activity index for assessing animal health and welfare.

CellRegNet: Point Annotation-Based Cell Detection in Histopathological Images via Density Map Regression.

Recent advances in deep learning have shown significant potential for accurate cell detection via density map regression using point annotations. However, existing deep learning models often struggle with multi-scale feature extraction and integration in complex histopathological images. Moreover, in multi-class cell detection scenarios, current density map regression methods typically predict each cell type independently, failing to consider the spatial distribution priors of different cell types. To address these challenges, we propose CellRegNet, a novel deep learning model for cell detection using point annotations. CellRegNet integrates a hybrid CNN/Transformer architecture with innovative feature refinement and selection mechanisms, addressing the need for effective multi-scale feature extraction and integration. Additionally, we introduce a contrastive regularization loss that models the mutual exclusiveness prior in multi-class cell detection cases. Extensive experiments on three histopathological image datasets demonstrate that CellRegNet outperforms existing state-of-the-art methods for cell detection using point annotations, with F1-scores of 86.38% on BCData (breast cancer), 85.56% on EndoNuke (endometrial tissue) and 93.90% on MBM (bone marrow cells), respectively. These results highlight CellRegNet's potential to enhance the accuracy and reliability of cell detection in digital pathology.

Comparative Analysis of YOLOv8 and YOLOv10 in Vehicle Detection: Performance Metrics and Model Efficacy

Accurate vehicle detection is crucial for the advancement of intelligent transportation systems, including autonomous driving and traffic monitoring. This paper presents a comparative analysis of two advanced deep learning models—YOLOv8 and YOLOv10—focusing on their efficacy in vehicle detection across multiple classes such as bicycles, buses, cars, motorcycles, and trucks. Using a range of performance metrics, including precision, recall, F1 score, and detailed confusion matrices, we evaluate the performance characteristics of each model.The findings reveal that YOLOv10 generally outperformed YOLOv8, particularly in detecting smaller and more complex vehicles like bicycles and trucks, which can be attributed to its architectural enhancements. Conversely, YOLOv8 showed a slight advantage in car detection, underscoring subtle differences in feature processing between the models. The performance for detecting buses and motorcycles was comparable, indicating robust features in both YOLO versions. This research contributes to the field by delineating the strengths and limitations of these models and providing insights into their practical applications in real-world scenarios. It enhances understanding of how different YOLO architectures can be optimized for specific vehicle detection tasks, thus supporting the development of more efficient and precise detection systems.

Advanced Forecasting of Drought Zones in Canada Using Deep Learning and CMIP6 Projections

This study addresses the critical issue of drought zoning in Canada using advanced deep learning techniques. Drought, exacerbated by climate change, significantly affects ecosystems, agriculture, and water resources. Canadian Drought Monitor (CDM) data provided by the Canadian government and ERA5-Land daily data were utilized to generate a comprehensive time series of mean monthly precipitation and air temperature for 199 sample locations in Canada from 1979 to 2023. These data were processed in the Google Earth Engine (GEE) environment and used to develop a Convolutional Neural Network (CNN) model to estimate CDM values, thereby filling gaps in historical drought data. The CanESM5 climate model, as assessed in the IPCC Sixth Assessment Report, was employed under four climate change scenarios to predict future drought conditions. Our CNN model forecasts CDM values up to 2100, enabling accurate drought zoning. The results reveal significant trends in temperature changes, indicating areas most vulnerable to future droughts, while precipitation shows a slow increasing trend. Our analysis indicates that under extreme climate scenarios, certain regions may experience a significant increase in the frequency and severity of droughts, necessitating proactive planning and mitigation strategies. These findings are critical for policymakers and stakeholders in designing effective drought management and adaptation programs.

Semantic segmentation for plant leaf disease classification and damage detection: A deep learning approach

Agriculture sustains the livelihoods of a significant portion of India's rural population, yet challenges persist in manual practices and disease management. To address these issues, this paper presents an automated plant leaf damage detection and disease identification system leveraging advanced deep learning techniques. The proposed method consists of six stages: first, utilizing YOLOv8 for region of interest identification from drone images; second, employing DeepLabV3+ for background removal and facilitating disease classification; third, implementing a CNN model for accurate disease classification achieving high training and validation accuracies (96.97 % and 92.89 %, respectively); fourth, utilizing UNet semantic segmentation for precise damage detection at a pixel level with an evaluation accuracy of 99 %; fifth, evaluating disease severity; and sixth, suggesting tailored remedies based on disease type and damage state. Experimental analysis using the Plant Village dataset demonstrates the effectiveness of the proposed method in detecting various defects in plants such as apple, tomato, and corn. This automated approach holds promise for enhancing agricultural productivity and disease management in India and beyond.