Generalization Capability Of Model Research Articles

Prediction of Protein–Protein Interactions Based on Integrating Deep Learning and Feature Fusion

Understanding protein–protein interactions (PPIs) helps to identify protein functions and develop other important applications such as drug preparation and protein–disease relationship identification. Deep-learning-based approaches are being intensely researched for PPI determination to reduce the cost and time of previous testing methods. In this work, we integrate deep learning with feature fusion, harnessing the strengths of both approaches, handcrafted features, and protein sequence embedding. The accuracies of the proposed model using five-fold cross-validation on Yeast core and Human datasets are 96.34% and 99.30%, respectively. In the task of predicting interactions in important PPI networks, our model correctly predicted all interactions in one-core, Wnt-related, and cancer-specific networks. The experimental results on cross-species datasets, including Caenorhabditis elegans, Helicobacter pylori, Homo sapiens, Mus musculus, and Escherichia coli, also show that our feature fusion method helps increase the generalization capability of the PPI prediction model.

Research on bearing fault diagnosis based on novel MRSVD-CWT and improved CNN-LSTM

Abstract As a critical component in mechanical equipment, rolling bearings play a vital role in industrial production. Effective bearing fault diagnosis provides a more reliable guarantee for the safe operation of the industrial output. Traditional data-driven bearing fault diagnosis methods often have problems such as insufficient fault feature extraction and poor model generalization capabilities, resulting in reduced diagnostic accuracy. To solve these problems and significantly improve the diagnosis accuracy, this paper proposes a novel fault diagnosis method based on multi-resolution singular value decomposition (MRSVD), continuous wavelet transform (CWT), improved convolutional neural network (CNN) enhanced by convolutional block attention module (CBAM), and long short-term memory (LSTM). Through MRSVD, the vibration signal is decomposed layer by layer into multiple denoised signals, thus signal noise can be eliminated to the greatest extent to gain the optimal denoised signals; then through CWT, the optimal denoised signals are converted into two-dimensional time-frequency images so that the local and global characteristic information can be fully captured. Finally, through improved CNN-LSTM, feature extraction is greatly enhanced, resulting in high accuracy of fault diagnosis. Lots of experiments are organized to test the performance, and the experimental results show that the proposed method on various datasets has better diagnosis accuracy and generalization ability under different working conditions than other methods.

Challenges in multi-centric generalization: phase and step recognition in Roux-en-Y gastric bypass surgery.

Most studies on surgical activity recognition utilizing artificial intelligence (AI) have focused mainly on recognizing one type of activity from small and mono-centric surgical video datasets. It remains speculative whether those models would generalize to other centers. In this work, we introduce a large multi-centric multi-activity dataset consisting of 140 surgical videos (MultiBypass140) of laparoscopic Roux-en-Y gastric bypass (LRYGB) surgeries performed at two medical centers, i.e., the University Hospital of Strasbourg, France (StrasBypass70) and Inselspital, Bern University Hospital, Switzerland (BernBypass70). The dataset has been fully annotated with phases and steps by two board-certified surgeons. Furthermore, we assess the generalizability and benchmark different deep learning models for the task of phase and step recognition in 7 experimental studies: (1) Training and evaluation on BernBypass70; (2) Training and evaluation on StrasBypass70; (3) Training and evaluation on the joint MultiBypass140 dataset; (4) Training on BernBypass70, evaluation on StrasBypass70; (5) Training on StrasBypass70, evaluation on BernBypass70; Training on MultiBypass140, (6) evaluation on BernBypass70 and (7) evaluation on StrasBypass70. The model's performance is markedly influenced by the training data. The worst results were obtained in experiments (4) and (5) confirming the limited generalization capabilities of models trained on mono-centric data. The use of multi-centric training data, experiments (6) and (7), improves the generalization capabilities of the models, bringing them beyond the level of independent mono-centric training and validation (experiments (1) and (2)). MultiBypass140 shows considerable variation in surgical technique and workflow of LRYGB procedures between centers. Therefore, generalization experiments demonstrate a remarkable difference in model performance. These results highlight the importance of multi-centric datasets for AI model generalization to account for variance in surgical technique and workflows. The dataset and code are publicly available at https://github.com/CAMMA-public/MultiBypass140.

SURVEY ON WBC CLASSIFICATION AND DETECTION USING YOLO

Our heart pumps red blood constantly, carrying waste products, hormones, nutrition, oxygen, and other essential health markers. Blood cells in this fluid are made up of many components, including platelets, RBCs, and WBCs. Neutrophils, Lymphocytes, Monocytes, Eosinophils, and Basophils are further subdivided into WBCs. To precisely identify and treat numerous kinds of blood-related illnesses, the detection and analysis of these blood cells is necessary for advanced diagnostics in medicine. Convolutional neural networks (CNNs), for example, are highly relied upon in modern medical diagnostics for the examination of blood cells. The precision and potency of blood cell identification procedures have been greatly improved by CNNs and Additional approaches to deep learning like RNNs and YOLO. The persistence of this research is to list challenges in White Blood cell detection and classification problem. Deep learning-based blood cell identification and classification approaches based on Single Stage approaches. Single Stage Approach involve You only look Once (YOLO). It explores the various architectures for feature selection and classification starting from YoloV1 toV5 and it’s various modifications. In this paper would be looking at things like network parameters, training procedure, and transfer learning application for each variation. It also assesses model complexity, accuracy, performance, generalization, computational efficiency, and transfer learning capabilities, in addition to variables like data amount and source.

Integrative BNN-LHS Surrogate Modeling and Thermo-Mechanical-EM Analysis for Enhanced Characterization of High-Frequency Low-Pass Filters in COMSOL.

This paper pioneers a novel approach in electromagnetic (EM) system analysis by synergistically combining Bayesian Neural Networks (BNNs) informed by Latin Hypercube Sampling (LHS) with advanced thermal-mechanical surrogate modeling within COMSOL simulations for high-frequency low-pass filter modeling. Our methodology transcends traditional EM characterization by integrating physical dimension variability, thermal effects, mechanical deformation, and real-world operational conditions, thereby achieving a significant leap in predictive modeling fidelity. Through rigorous evaluation using Mean Squared Error (MSE), Maximum Learning Error (MLE), and Maximum Test Error (MTE) metrics, as well as comprehensive validation on unseen data, the model's robustness and generalization capability is demonstrated. This research challenges conventional methods, offering a nuanced understanding of multiphysical phenomena to enhance reliability and resilience in electronic component design and optimization. The integration of thermal variables alongside dimensional parameters marks a novel paradigm in filter performance analysis, significantly improving simulation accuracy. Our findings not only contribute to the body of knowledge in EM diagnostics and complex-environment analysis but also pave the way for future investigations into the fusion of machine learning with computational physics, promising transformative impacts across various applications, from telecommunications to medical devices.

Forecasting cryptocurrency's buy signal with a bagged tree learning approach to enhance purchase decisions.

The cryptocurrency market is captivating the attention of both retail and institutional investors. While this highly volatile market offers investors substantial profit opportunities, it also entails risks due to its sensitivity to speculative news and the erratic behavior of major investors, both of which can provoke unexpected price fluctuations. In this study, we contend that extreme and sudden price changes and atypical patterns might compromise the performance of technical signals utilized as the basis for feature extraction in a machine learning-based trading system by either augmenting or diminishing the model's generalization capability. To address this issue, this research uses a bagged tree (BT) model to forecast the buy signal for the cryptocurrency market. To achieve this, traders must acquire knowledge about the cryptocurrency market and modify their strategies accordingly. To make an informed decision, we depended on the most prevalently utilized oscillators, namely, the buy signal in the cryptocurrency market, comprising the Relative Strength Index (RSI), Bollinger Bands (BB), and the Moving Average Convergence/Divergence (MACD) indicator. Also, the research evaluates how accurately a model can predict the performance of different cryptocurrencies such as Bitcoin (BTC), Ethereum (ETH), Cardano (ADA), and Binance Coin (BNB). Furthermore, the efficacy of the most popular machine learning model in precisely forecasting outcomes within the cryptocurrency market is examined. Notably, predicting buy signal values using a BT model provides promising results.

Application of Artificial Intelligence-based Technology in College Archives Management

This study looks into the use of artificial intelligence (AI)-based technologies, notably Convolutional Neural Network (CNN) models and data augmentation approaches, in college archives management. With the rapid expansion of digital information, archive organizations have enormous issues in maintaining, organizing, and giving access to their vast collections of historical documents and scholarly resources. AI solutions present possible solutions to these difficulties by automating labor-intensive operations and improving the efficiency and accuracy of preservation workflows. In this study, they investigate the transformational potential of CNN models in digitization activities, utilizing their visual data processing skills to automate and enhance the digitization of visual resources within college archives. In addition, they look into how data augmentation techniques might improve the resilience and generalization capabilities of AI models, reducing biases and improving performance across a variety of archive collections. Using empirical research, theoretical analysis, and case studies, they present insights into the opportunities and obstacles of implementing AI-based technology in college archive management. Furthermore, they discuss ethical implications for the deployment of AI solutions, emphasizing the significance of transparency, accountability, and stakeholder participation. This study aims to inform archival practitioners, technologists, and stakeholders about the effective and responsible integration of AI technologies in cultural heritage preservation and promotion within academic institutions by elucidating the practical applications and implications of CNN models and data augmentation techniques in archival contexts.

Topology optimization based on improved DoubleU-Net using four boundary condition datasets

Topology optimization is extensively employed in engineering design. However, applying deep learning to solve topology optimization problems faces challenges owing to limited training data and model adaptability to various boundary conditions. To address these challenges, this research has employed a solid isotropic material with a penalization method to produce a dataset of 400,000 2D samples covering four boundary conditions. Each boundary condition includes two resolution datasets. Moreover, an improved DoubleU-Net model is proposed for real-time topology optimization, ensuring high-accuracy predictions. On the test dataset, the average Intersection over Union (IoU) accuracies of the proposed model for the four boundary conditions above are 93.26%, 96.71%, 96.35% and 97.38%. This research also investigated the impact of different structural datasets on the model's generalization capability. The experimental results indicate that the model trained on random structures exhibits excellent testing performance and demonstrates robust generalization capability.

YOLOv8-MU: An Improved YOLOv8 Underwater Detector Based on a Large Kernel Block and a Multi-Branch Reparameterization Module.

Underwater visual detection technology is crucial for marine exploration and monitoring. Given the growing demand for accurate underwater target recognition, this study introduces an innovative architecture, YOLOv8-MU, which significantly enhances the detection accuracy. This model incorporates the large kernel block (LarK block) from UniRepLKNet to optimize the backbone network, achieving a broader receptive field without increasing the model's depth. Additionally, the integration of C2fSTR, which combines the Swin transformer with the C2f module, and the SPPFCSPC_EMA module, which blends Cross-Stage Partial Fast Spatial Pyramid Pooling (SPPFCSPC) with attention mechanisms, notably improves the detection accuracy and robustness for various biological targets. A fusion block from DAMO-YOLO further enhances the multi-scale feature extraction capabilities in the model's neck. Moreover, the adoption of the MPDIoU loss function, designed around the vertex distance, effectively addresses the challenges of localization accuracy and boundary clarity in underwater organism detection. The experimental results on the URPC2019 dataset indicate that YOLOv8-MU achieves an mAP@0.5 of 78.4%, showing an improvement of 4.0% over the original YOLOv8 model. Additionally, on the URPC2020 dataset, it achieves 80.9%, and, on the Aquarium dataset, it reaches 75.5%, surpassing other models, including YOLOv5 and YOLOv8n, thus confirming the wide applicability and generalization capabilities of our proposed improved model architecture. Furthermore, an evaluation on the improved URPC2019 dataset demonstrates leading performance (SOTA), with an mAP@0.5 of 88.1%, further verifying its superiority on this dataset. These results highlight the model's broad applicability and generalization capabilities across various underwater datasets.

MFNet: Meta-learning based on frequency-space mix for MRI segmentation in nasopharyngeal carcinoma.

Deep learning techniques have been applied to medical image segmentation and demonstrated expert-level performance. Due to the poor generalization abilities of the models in the deployment in different centres, common solutions, such as transfer learning and domain adaptation techniques, have been proposed to mitigate this issue. However, these solutions necessitate retraining the models with target domain data and annotations, which limits their deployment in clinical settings in unseen domains. We evaluated the performance of domain generalization methods on the task of MRI segmentation of nasopharyngeal carcinoma (NPC) by collecting a new dataset of 321 patients with manually annotated MRIs from two hospitals. We transformed the modalities of MRI, including T1WI, T2WI and CE-T1WI, from the spatial domain to the frequency domain using Fourier transform. To address the bottleneck of domain generalization in MRI segmentation of NPC, we propose a meta-learning approach based on frequency domain feature mixing. We evaluated the performance of MFNet against existing techniques for generalizing NPC segmentation in terms of Dice and MIoU. Our method evidently outperforms the baseline in handling the generalization of NPC segmentation. The MF-Net clearly demonstrates its effectiveness for generalizing NPC MRI segmentation to unseen domains (Dice = 67.59%, MIoU = 75.74% T1W1). MFNet enhances the model's generalization capabilities by incorporating mixed-feature meta-learning. Our approach offers a novel perspective to tackle the domain generalization problem in the field of medical imaging by effectively exploiting the unique characteristics of medical images.

ContourTL-Net: Contour-Based Transfer Learning Algorithm for Early-Stage Brain Tumor Detection.

Brain tumors are critical neurological ailments caused by uncontrolled cell growth in the brain or skull, often leading to death. An increasing patient longevity rate requires prompt detection; however, the complexities of brain tissue make early diagnosis challenging. Hence, automated tools are necessary to aid healthcare professionals. This study is particularly aimed at improving the efficacy of computerized brain tumor detection in a clinical setting through a deep learning model. Hence, a novel thresholding-based MRI image segmentation approach with a transfer learning model based on contour (ContourTL-Net) is suggested to facilitate the clinical detection of brain malignancies at an initial phase. The model utilizes contour-based analysis, which is critical for object detection, precise segmentation, and capturing subtle variations in tumor morphology. The model employs a VGG-16 architecture priorly trained on the "ImageNet" collection for feature extraction and categorization. The model is designed to utilize its ten nontrainable and three trainable convolutional layers and three dropout layers. The proposed ContourTL-Net model is evaluated on two benchmark datasets in four ways, among which an unseen case is considered as the clinical aspect. Validating a deep learning model on unseen data is crucial to determine the model's generalization capability, domain adaptation, robustness, and real-world applicability. Here, the presented model's outcomes demonstrate a highly accurate classification of the unseen data, achieving a perfect sensitivity and negative predictive value (NPV) of 100%, 98.60% specificity, 99.12% precision, 99.56% F1-score, and 99.46% accuracy. Additionally, the outcomes of the suggested model are compared with state-of-the-art methodologies to further enhance its effectiveness. The proposed solution outperforms the existing solutions in both seen and unseen data, with the potential to significantly improve brain tumor detection efficiency and accuracy, leading to earlier diagnoses and improved patient outcomes.

Masked autoencoders with generalizable self-distillation for skin lesion segmentation.

In the field of skin lesion image segmentation, accurate identification and partitioning of diseased regions is of vital importance for in-depth analysis of skin cancer. Self-supervised learning, i.e., MAE, has emerged as a potent force in the medical imaging domain, which autonomously learns and extracts latent features from unlabeled data, thereby yielding pre-trained models that greatly assist downstream tasks. To encourage pre-trained models to more comprehensively learn the global structural and local detail information inherent in dermoscopy images, we introduce a Teacher-Student architecture, named TEDMAE, by incorporating a self-distillation mechanism, it learns holistic image feature information to improve the generalizable global knowledge learning of the student MAE model. To make the image features learned by the model suitable for unknown test images, two optimization strategies are, Exterior Conversion Augmentation (EC) utilizes random convolutional kernels and linear interpolation to effectively transform the input image into one with the same shape but altered intensities and textures, while Dynamic Feature Generation (DF) employs a nonlinear attention mechanism for feature merging, enhancing the expressive power of the features, are proposed to enhance the generalizability of global features learned by the teacher model, thereby improving the overall generalization capability of the pre-trained models. Experimental results from the three public skin disease datasets, ISIC2019, ISIC2017, and PH indicate that our proposed TEDMAE method outperforms several similar approaches. Specifically, TEDMAE demonstrated optimal segmentation and generalization performance on the ISIC2017 and PH datasets, with Dice scores reaching 82.1% and 91.2%, respectively. The best Jaccard values were 72.6% and 84.5%, while the optimal HD95% values were 13.0% and 8.9%, respectively.

A method for quantifying the generalization capabilities of generative models for solving Ising models

For Ising models with complex energy landscapes, whether the ground state can be found by neural networks depends heavily on the Hamming distance between the training datasets and the ground state. Despite the fact that various recently proposed generative models have shown good performance in solving Ising models, there is no adequate discussion on how to quantify their generalization capabilities. Here we design a Hamming distance regularizer in the framework of a class of generative models, variational autoregressive networks (VANs), to quantify the generalization capabilities of various network architectures combined with VAN. The regularizer can control the size of the overlaps between the ground state and the training datasets generated by networks, which, together with the success rates of finding the ground state, form a quantitative metric to quantify their generalization capabilities. We conduct numerical experiments on several prototypical network architectures combined with VAN, including feed-forward neural networks, recurrent neural networks, and graph neural networks, to quantify their generalization capabilities when solving Ising models. Moreover, considering the fact that the quantification of the generalization capabilities of networks on small-scale problems can be used to predict their relative performance on large-scale problems, our method is of great significance for assisting in the Neural Architecture Search field of searching for the optimal network architectures when solving large-scale Ising models.

High generalization identification method based on MI-SI distributed optical fiber sensor and video signals

A pattern recognition method for intrusion signals is introduced, leveraging distributed optical fiber sensors and convolutional neural networks. To enhance the model's generalization capabilities, two strategies are employed. Firstly, the sensor structure is refined, incorporating the strengths of both the Sagnac and Michelson interferometers to create a hybrid design with a wide frequency response. Comparative experiments confirm that this exceptional frequency response significantly boosts the model's generalization power. Secondly, to address the limitations of distributed fiber optic sensors, a multimodal approach is proposed, integrating video frame difference and fiber optic pattern recognition. The trained convolutional neural network model exhibits low computational overhead, robust generalization, and practical applicability. When applied to a dataset collected in diverse scenarios not included in the model training, the model achieved a CNR of 0% and a FPR of only 0.26% for the intrusion behavior of climbing fences.

PPA-SAM: Plug-and-Play Adversarial Segment Anything Model for 3D Tooth Segmentation

In Cone Beam Computed Tomography (CBCT) images, accurate tooth segmentation is crucial for oral health, providing essential guidance for dental procedures such as implant placement and difficult tooth extractions (impactions). However, due to the lack of a substantial amount of dental data and the complexity of tooth morphology in CBCT images, the task of tooth segmentation faces significant challenges. This may lead to issues such as overfitting and training instability in existing algorithms, resulting in poor model generalization. Ultimately, this may impact the accuracy of segmentation results and could even provide incorrect diagnostic and treatment information. In response to these challenges, we introduce PPA-SAM, an innovative dual-encoder segmentation network that merges the currently popular Segment Anything Model (SAM) with the 3D medical segmentation network, VNet. Through the use of adapters, we achieve parameter reuse and fine-tuning, enhancing the model’s adaptability to specific CBCT datasets. Simultaneously, we utilize a three-layer convolutional network as both a discriminator and a generator for adversarial training. The PPA-SAM model seamlessly integrates the high-precision segmentation performance of convolutional networks with the outstanding generalization capabilities of SAM models, achieving more accurate and robust three-dimensional tooth segmentation in CBCT images. Evaluation of a small CBCT dataset demonstrates that PPA-SAM outperforms other networks in terms of accuracy and robustness, providing a reliable and efficient solution for three-dimensional tooth segmentation in CBCT images. This research has a positive impact on the management of dentofacial conditions from oral implantology to orthognathic surgery, offering dependable technological support for future oral diagnostics and treatment planning.

A novel hybrid ensemble model for mineral prospectivity prediction: A case study in the Malipo W-Sn mineral district, Yunnan Province, China

The utilization of machine learning and data-driven approaches for mineral prospectivity modeling has become increasingly important in supporting regional mineral exploration. However, supervised learning methods have encountered several common challenges, such as overfitting and underfitting, in the process of mineral prospectivity modeling, particularly due to imbalanced sample sizes used for model training. This imbalance could limit the generalization capability of models and result in inadequate predictive performance of individual models. The present study introduced a novel hybrid ensemble model (HEM) that integrates various learner methods, including generative adversarial network (GAN) and transfer learning techniques, to effectively address the issues above. These intelligent prediction techniques were applied to analyze the geological, geochemical, and geophysical dataset in order to delineate potential areas for W-Sn mineralization in Malipo County, Yunnan Province, China. Meanwhile, some well-established supervised learning methods such as Convolutional Neural Networks (CNN), Boosting-based model (e.g., LightGBM), and Bagging-based models, which demonstrated exceptional performance in dealing with classification issues were also integrated into the HEM. The weights for integrating these models were determined automatically using the equidistant-sampling technique. The superior performance of the HEM in identifying known deposits made it an efficient tool for mineral prediction mapping. Additionally, this model also offered high flexibility and scalability by allowing adjustments to the types and quantities of base learners based on data characteristics.

CAT-DTI: cross-attention and Transformer network with domain adaptation for drug-target interaction prediction

Accurate and efficient prediction of drug-target interaction (DTI) is critical to advance drug development and reduce the cost of drug discovery. Recently, the employment of deep learning methods has enhanced DTI prediction precision and efficacy, but it still encounters several challenges. The first challenge lies in the efficient learning of drug and protein feature representations alongside their interaction features to enhance DTI prediction. Another important challenge is to improve the generalization capability of the DTI model within real-world scenarios. To address these challenges, we propose CAT-DTI, a model based on cross-attention and Transformer, possessing domain adaptation capability. CAT-DTI effectively captures the drug-target interactions while adapting to out-of-distribution data. Specifically, we use a convolution neural network combined with a Transformer to encode the distance relationship between amino acids within protein sequences and employ a cross-attention module to capture the drug-target interaction features. Generalization to new DTI prediction scenarios is achieved by leveraging a conditional domain adversarial network, aligning DTI representations under diverse distributions. Experimental results within in-domain and cross-domain scenarios demonstrate that CAT-DTI model overall improves DTI prediction performance compared with previous methods.

Bridging the gap between crop breeding and GeoAI: Soybean yield prediction from multispectral UAV images with transfer learning

Despite significant progress has been made towards crop yield prediction with remote sensing, there exist knowledge gaps on (1) the impacts of temporal resolution of imaging frequencies on yield prediction, (2) transferability of the models among different genotypes and test sites, and (3) translation of these research developments to crop breeding that benefit farmers. Existing research predominantly provides an on-site perspective, frequently missing the complexities of real-world applications. The objectives of this paper are to investigate the transferability and generalization capabilities of yield prediction models for crop breeding across test sites located in North and South Americas. Toward that goal, we tested different machine learning techniques including Random Forest Regressor (RF), Gradient Boosting Regression (GB), and Deep Neural Networks (DNN) for soybean yield prediction with experiments conducted in different climate and growth conditions. A novel transfer learning approach was proposed for genotype selection and categorizing soybean yield for screening high-yield varieties. Furthermore, we studied the effect of temporal resolution on yield prediction, focusing on the critical development stages and optimal aerial survey frequencies for precise yield prediction using large 31,404 sample data. Results demonstrated that the combined dataset of Argentina and United States representing different climate regimes provided the highest performance with an R2 of 0.76 using RF and GB algorithms. The classification approach was proven to be most useful for crop breeding as demonstrated by accurately identifying the high-yielding genotypes. Increasing temporal sampling of key phenological stages significantly improved yield prediction. Although transfer learning yielded promising outcomes across trials within Argentina the efficacy of transferring models from Argentina to the United States was limited, attributed to significant seasonal and climate variations. This study pioneered the use of transfer learning for model adaptability in real-world breeding scenarios, training and transferring models within the South and North Americas, providing actionable insights and strategies for the breeding community, aiming to facilitate improved decision-making for agricultural productivity.

Text data augmentation and pre-trained Language Model for enhancing text classification of low-resource languages.

In the domain of natural language processing (NLP), the development and success of advanced language models are predominantly anchored in the richness of available linguistic resources. Languages such as Azerbaijani, which is classified as a low-resource, often face challenges arising from limited labeled datasets, consequently hindering effective model training. The primary objective of this study was to enhance the effectiveness and generalization capabilities of news text classification models using text augmentation techniques. In this study, we solve the problem of working with low-resource languages using translations using the Facebook mBart50 model, as well as the Google Translate API and a combination of mBart50 and Google Translate thus expanding the capabilities when working with text. The experimental outcomes reveal a promising uptick in classification performance when models are trained on the augmented dataset compared with their counterparts using the original data. This investigation underscores the immense potential of combined data augmentation strategies to bolster the NLP capabilities of underrepresented languages. As a result of our research, we have published our labeled text classification dataset and pre-trained RoBERTa model for the Azerbaijani language.

Patch-Level Feature Selection for Thoracic Disease Classification by Chest X-ray Images Using Information Bottleneck.

Chest X-ray (CXR) examination serves as a widely employed clinical test in medical diagnostics. Many studied have tried to apply artificial intelligence (AI) programs to analyze CXR images. Despite numerous positive outcomes, assessing the applicability of AI models for comprehensive diagnostic support remains a formidable challenge. We observed that, even when AI models exhibit high accuracy on one dataset, their performance may deteriorate when tested on another. To address this issue, we propose incorporating a variational information bottleneck (VIB) at the patch level to enhance the generalizability of diagnostic support models. The VIB introduces a probabilistic model aimed at approximating the posterior distribution of latent variables given input data, thereby enhancing the model's generalization capabilities on unseen data. Unlike the conventional VIB approaches that flatten features and use a re-parameterization trick to sample a new latent feature, our method applies the trick to 2D feature maps. This design allows only important pixels to respond, and the model will select important patches in an image. Moreover, the proposed patch-level VIB seamlessly integrates with various convolutional neural networks, offering a versatile solution to improve performance. Experimental results illustrate enhanced accuracy in standard experiment settings. In addition, the method shows robust improvement when training and testing on different datasets.