Explain and improve: LRP-inference fine-tuning for image captioning models

Abstract: The first part of the motivations behind the demand for explainable and interpretable models, emphasizing the ethical, legal, and practical implications of deploying black-box models in critical domains. The discussion extends to the societal impact of decisions made by these models and the importance of building trust among end-users. The second part explores the existing techniques and methodologies designed to enhance the explain ability and interpretability of ML models. From traditional linear models to complex deep neural networks, delve into methods. Special attention is given to recent advancements, including state-of-the-art attention mechanisms and layer-wise relevance propagation in neural networks. Search significance: The primary goal of any machine learning model is to demonstrate high ability by effectively capturing patterns, relationships, and trends within the training data and generalizing well to new, unseen data. A model with high ability is capable of making accurate predictions, which is crucial for its practical utility and effectiveness in various applications. Interpretability is essential for building trust and acceptance of machine learning models. An interpretable model allows stakeholders to comprehend why a particular prediction was made, which is crucial for model deployment, regulatory compliance, and ethical considerations. Method: the approach involving a weighted sum essentially involves performing multiplication, while subtraction is employed for addition and sorting. When considering candidate keywords, we’ve previously explored how they are generated and presented. The weighted sum of a fourdimensional feature constitutes the vector, which, however, is altered in the course of the process. The necessity for weights arises, as the four characteristics possess different parsing capabilities. This encompasses both keywords and keywords. The greater the ability to differentiate, the more effective the manual identification process can be. In reality, manually performing the identification of a weight vector for the domain proves to be excessively burdensome due to its complexity. Result: From the result Random Forest is in 1st rank whereas DNN 5 Layers is in lowest rank Keywords: machine learning; explain ability; interpretability; fairness; sensitivity; black-box

Pneumonia remains a global health issue, creating the need for accurate detection methods for effective treatment. Deep learning models like ResNet50 show promise in detecting pneumonia from chest X-rays; however, their black-box nature limits the transparency, which fails to meet that needed for clinical trust. This study aims to improve model interpretability by comparing four interpretability techniques, which are Layer-wise Relevance Propagation (LRP), Adversarial Training, Class Activation Maps (CAMs), and the Spatial Attention Mechanism, and determining which fits best the model, enhancing its transparency with minimal impact on its performance. Each technique was evaluated for its impact on the accuracy, sensitivity, specificity, AUC-ROC, Mean Relevance Score (MRS), and a calculated trade-off score that balances interpretability and performance. The results indicate that LRP was the most effective in enhancing interpretability, achieving high scores across all metrics without sacrificing diagnostic accuracy. The model achieved 0.91 accuracy and 0.85 interpretability (MRS), demonstrating its potential for clinical integration. In contrast, Adversarial Training, CAMs, and the Spatial Attention Mechanism showed trade-offs between interpretability and performance, each highlighting unique image features but with some impact on specificity and accuracy.

ClathPLM: Deep multi-view feature extraction with CNN and attention enhances clathrin protein identification.

The advancement of artificial intelligence (AI) tools utilizing electroencephalography (EEG) for diagnosing major depressive disorder (MDD) has shown significant progress. However, the practical implementation of these tools is often impeded by the large amount of EEG data required for training AI models and the lack of explanations for the MDD diagnoses. This study aims to develop an interpretable deep-learning-based computer-aided diagnostic system for diagnosing male MDD patients using explainable AI (XAI) algorithms. The CAD system was designed to facilitate the diagnostic process by using a reduced number of EEG channels and data length while enhancing understanding of the neurophysiological characteristics of male MDD. Resting-state EEG data were collected from 40 male MDD patients (20–63 years) and 41 gender-matched healthy controls (HCs, 19–61 years). A shallow convolutional neural network (CNN; Shallow ConvNet) model was utilized to distinguish between MDD patients and HCs. Relevance scores were extracted by the layer-wise relevance propagation (LRP) method, integrated with the Shallow ConvNet, to interpret the outcomes of the deep-learning-based CAD system. Additionally, changes in diagnostic performance were assessed by progressively reducing the number of channels using an LRP-based channel selection method, as well as EEG data length. Our XAI-based CAD system showed a high diagnostic performance of 100% when using the whole 62 channels with 180-s EEG data. A relatively high diagnostic performance over 90% was retained with only five channels with 60-s EEG data. Neurophysiologically meaningful brain areas, such as fronto-central, centro-parietal, and occipital areas, also revealed significant differences in relevance scores extracted by the LRP-method between the two groups. This study successfully developed a high performance and practical XAI-based CAD system for male MDD patients. Our developed CAD system not only achieves high diagnostic accuracy but also provides meaningful neurophysiological biomarkers for male MDD patients.

Synthetic aperture radar (SAR) automatic target recognition (ATR) models based on deep neural networks (DNNs) are susceptible to adversarial attacks. In this study, we proposed an SAR black-box local adversarial attack algorithm named attention heat map- based black-box local adversarial attack (AH-BLAA). First, we designed an attention heat map extraction module combined with the layer-wise relevance propagation (LRP) algorithm to obtain the high concerning areas of the SAR-ATR models. Then, to gener- ate SAR adversarial attack examples, we designed a perturbation generator module, introducing the structural dissimilarity (DSSIM) metric in the loss function to limit image distortion and the dif- ferential evolution (DE) algorithm to search for optimal perturba- tions. Experimental results on the MSTAR and FUSAR-Ship datasets showed that compared with existing adversarial attack algorithms, the attack success rate of the AH-BLAA algorithm increased by 0.63% to 33.59% and 1.05% to 17.65%, respectively. Moreover, the low- est perturbation ratios reached 0.23% and 0.13%, respectively.

BackgroundDeep learning has shown promising results regarding Alzheimer’s disease (AD) studies. In the meantime, heatmap methods have emerged as a popular tool to visualize deep learning models, enhancing the transparency and explainability of deep learning. However, when no ground truth is available, the heatmap methods are particularly difficult to trust. In the neuroimaging field, meta‐analysis is often conducted to reduce heterogeneity and reveal the true positive. Thus, in this work, we aim at validating heatmap methods by quantifying the overlap of a ground truth map provided by a large meta‐analysis and heatmaps derived by three methods, the Layer‐wise Relevance Propagation (LRP) method, the Integrated Gradients (IG) method, and the Guided grad‐CAM (GGC) method from convolutional neural network achieving state‐of‐the‐art classification performance for AD classification on a sample of MRI scans from ADNI.Method502 ADNI participants were included in this study with 250 AD participants and 252 healthy controls. Those brain scans were trained by five different 3D CNN architectures. Three CNN heatmap methods were applied on trained models to produce a heatmap indicating which voxels were the most important when classifying AD and control participants. The structural MRI meta‐analysis map used as the ground truth summarized 77 neuroimaging studies reporting 773 locations in the brain affected by AD.ResultThe best CNN 5‐fold cross validated accuracy reached 87.25% for classifying AD and CN. The best Dice overlap measured for LRP, IG, and GGC, using this model was 0.502 for LRP heatmap, 0.55 for the IG heatmap, and 0.54 for the GGC heatmap. SVM activation patterns achieved a dice of 0.363.ConclusionWe found that the three heatmap methods capture brain regions that overlap well with the meta‐analysis map, and we observed the best overlap for the IG method. All three heatmap methods outperformed linear SVM models suggesting that the deep feature learnt by the most recent deep neural networks can produce more meaningful representations than linear and shallow models.

Autism spectrum disorder is a very common disorder. An early diagnosis of autism is essential for the prognosis of this disorder. The common diagnosis method utilizes behavioural cues of autistic children. Doctors require years of clinical training to acquire the ability to capture these behavioural cues (such as self–stimulatory behaviours). In recent years, the advancement of deep learning algorithms and hardware enabled the use of artificial intelligence technology to automatically capture self-stimulatory behaviours. Using this technique, the work efficacy of doctors can be improved. However, the field of self-stimulatory behaviours research still lacks large annotated data to train the model. Therefore, the application of unsupervised machine learning methods is adopted. Meanwhile, it is often difficult to obtain good classification results using unlabelled data, further research to train a model that can obtain good classification results and at the same time being practical will be valuable. Nevertheless, in the area of machine learning, the interpretability of the created model has to be vital as well. Hence, we have employed the Layer-wise Relevance Propagation (LRP) method to explain the proposed model. In this article, the major innovation is utilizing the temporal coherency between adjacent frames as free supervision and setting a global discriminative margin to extract slow-changing discriminative self-stimulatory behaviours features. Extensive evaluation of the extracted features has proven the effectiveness of those features. Firstly, the extracted features are classified by the k-means method to show the classification of self-stimulation behaviours in a completely unsupervised way. Then, the conditional entropy method is used to evaluate the effectiveness of features. Secondly, we have obtained the state-of-the-art results by combining the unsupervised TCDN method with optimised supervised learning methods (such as SVM, k-NN, Discriminant). These state-of-the-art results prove the effectiveness of the slow-changing discriminative self-stimulatory behaviours features.

An adaptive attention mechanism algorithm based on multiscale fusion is proposed to solve the nonuniform blind blur of image denoising algorithm for depth of field, camera wobble and object motion. Firstly, the convolutional neural network residual module (MRESNET) was established to extract the deep features of the image, and the image pyramid was used as input to the model. Channel attentional mechanism (SE) and spatial attentional mechanism fusion (CBAM) were inserted into the network to improve local and spatial image representation. This paper introduces Generative Antagonism Network (GAN), and because of the sparseness of noise image, sets up discriminator, separates fuzzy information from useful information in training iteration, and then distinguishes real image from de-noise image. Experimental results show that the proposed algorithm improves peak SNR (PSNR) and image similarity (SSIM) and can restore higher quality images.

This study proposes a multiheaded object detection algorithm referred to as MANet. The main purpose of the study is to integrate feature layers of different scales based on the attention mechanism and to enhance contextual connections. To achieve this, we first replaced the feed-forward base network of the single-shot detector with the ResNet–101 (inspired by the Deconvolutional Single-Shot Detector) and then applied linear interpolation and the attention mechanism. The information of the feature layers at different scales was fused to improve the accuracy of target detection. The primary contributions of this study are the propositions of (a) a fusion attention mechanism, and (b) a multiheaded attention fusion method. Our final MANet detector model effectively unifies the feature information among the feature layers at different scales, thus enabling it to detect objects with different sizes and with higher precision. We used the 512 × 512 input MANet (the backbone is ResNet–101) to obtain a mean accuracy of 82.7% based on the PASCAL visual object class 2007 test. These results demonstrated that our proposed method yielded better accuracy than those provided by the conventional Single-shot detector (SSD) and other advanced detectors.

The increasing sophistication and frequency of cyberattacks necessitate advanced and transparent decision-making frameworks for threat intelligence and risk assessment. While artificial intelligence (AI) has demonstrated significant potential in automating threat detection, predicting attack patterns, and prioritizing incident response, the opaque nature of many AI models particularly deep learning poses challenges in trust, interpretability, and regulatory compliance. This paper presents an Explainable AI (XAI)-driven framework for enhancing Cyber Threat Intelligence (CTI) and risk assessment, enabling security analysts to understand, validate, and act upon AI-generated insights with confidence. The proposed approach integrates state-of-the-art machine learning algorithms with model-agnostic and model-specific interpretability techniques, such as SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), Layer-wise Relevance Propagation (LRP), and attention-based visualization. The framework processes diverse CTI data sources including network traffic logs, vulnerability feeds, malware signatures, and open-source intelligence (OSINT) to identify malicious patterns, quantify risks, and generate actionable intelligence. By coupling predictive accuracy with transparency, the system not only improves detection and classification of advanced persistent threats (APTs), phishing campaigns, and zero-day exploits, but also facilitates compliance with emerging AI governance frameworks, such as the EU AI Act and NIST AI Risk Management Framework. Experimental evaluations on benchmark cybersecurity datasets demonstrate that the XAI-enabled models maintain high precision and recall while offering interpretable outputs that significantly reduce analyst decision latency and improve collaborative threat response. The paper further discusses the role of XAI in mitigating bias, enhancing accountability, and fostering human–machine teaming in security operations centers (SOCs). Practical implementation considerations including scalability, computational efficiency, and integration with existing Security Information and Event Management (SIEM) platforms are addressed, along with recommendations for future research in multi-modal XAI for cybersecurity. This research underscores the critical value of explainability in AI-driven CTI, ensuring that automated systems are not only effective but also transparent, trustworthy, and aligned with organizational and regulatory requirements for responsible AI deployment in cyber defense.

To identify various oil well working conditions more accurately and practically from massive image data collected by multiple measured information sources of sucker-rod pumping wells, this paper proposes a working condition recognition method with three key aspects: curvelet pooling optimization technology, multi-source attention mechanism fusion feature extraction technology, and multi-source semi-supervised classification deep learning. Specifically: (a) Curvelet pooling optimization technology. We introduce the second-generation curvelet transform into the ResNet-50 pooling layer and adopt a collaborative learning pooling strategy of low-frequency and high-frequency information from the raw data decomposed via curvelet transform instead of max-pooling. This enhances the neural network’s capability to capture detailed features of complex image data. (b) Multi-source attention mechanism fusion feature extraction technology. We selected two information sources: measured ground dynamometer cards and measured electrical power cards. The multi-head self-attention mechanism enables interactive complementarity between curvelet-decomposed image data from each information source, while achieving dynamic weighted fusion of the interactive complementary data via the adaptive attention mechanism. This process yields optimal global feature representations of multi-source fused data. (c) Multi-source semi-supervised classification deep learning. By integrating multi-source fused feature data with a semi-supervised classification algorithm based on the dual strategy of dynamic adjustment of pseudo-label confidence and self-adaptive class fairness regularization, the method leverages abundant multi-source unlabeled samples to improve model classification performance and generalization ability under limited labeled training samples. This further enhances the accuracy and practicality of condition recognition. Experimental data were collected from a high-pressure, low-permeability, thin oil reservoir block in an oilfield in China. Extensive experiments demonstrate that the proposed method efficiently processes measured information source data in the sucker-rod pumping production system, improves the performance of traditional deep learning frameworks, explores the intrinsic correlations among multiple measured information source data of oil wells, and utilizes massive unlabeled working condition data to enhance the working condition recognition effect and engineering practicability with a minimal number of labeled samples. Code is available at https://github.com/Yoick/AMMFFECP.

To better understand and evaluate the GNSS positioning performance, it is convenient to adopt corresponding measures to reduce the impact of errors on positioning. A GNSS positioning error estimation scheme based on an improved multi-layer perceptron model is proposed. The multi-head attention mechanism and gating operation are integrated into the multi-layer perceptron model to adaptively select and filter features, enhancing the model’s ability to understand input features. First, the original positioning error of the satellite is obtained through the Kalman filter positioning method. The data are then preprocessed to extract available features. Finally, the features are input into the constructed model for training and testing to obtain the estimated positioning error value. Two types of comparative experiments were completed. The performance of the presented model is evaluated by the root mean square error. Experimental results show that the proposed method performs well in terms of performance indicators, and has obvious advantages over other state-of-the-art methods. In particular, the root mean square error of the presented method in the first dataset is 0.239 m, which is 39.2% and 17% lower than the current state-of-the-art long short-term memory network and convolutional neural network, respectively. The presented method can provide higher-precision estimated values for studying the GNSS positioning error estimation problem.

To overcome the limitations of existing deep learning-based image recognition approaches for diagnosing the operating conditions of pumping wells–namely, slow processing speed and high parameter complexity–this paper proposes a lightweight diagnostic method based on a one-dimensional convolutional neural network (1D CNN). Unlike traditional image-based techniques, the proposed method eliminates the need to convert raw data into dynamometer card images, thereby reducing computational overhead and significantly decreasing the number of model parameters. The architecture integrates 1D CNN with a bidirectional long short-term memory (BiLSTM) network to effectively capture both local patterns and global temporal dependencies inherent in different operating states of pumping wells. Additionally, a multi-head attention (MHA) mechanism is incorporated to enhance the representation of critical time points, thereby improving the model's sensitivity to key features and its overall classification performance. Experiments conducted on an augmented dataset–generated via a simulation-based algorithm–demonstrate that the proposed 1DCNN-BiLSTM-MHA model achieves a classification accuracy of 99.22%. Experimental results demonstrate that the proposed model possesses strong feature learning capabilities. Comparative analysis with several baseline models highlights its superior diagnostic performance and reliability, along with reduced parameter count and improved computational speed.

COVID-19 fake news detection: A hybrid CNN-BiLSTM-AM model

Piwi interacting RNA (piRNA) is a type of small non coding RNA with a length of 24-32 nucleotides, mainly expressed in germ cells. Its abnormal expression is closely related to various diseases such as cancer and neurodegenerative diseases. Although biological experiments are the gold standard for identifying the association between piRNA and disease (PDA), their high cost and long cycle limit research progress. Therefore, computational models have become an important tool for assisting in predicting PDA. However, existing computational methods generally suffer from issues such as insufficient feature extraction and imbalanced data. This article proposes a prediction model based on the fusion of graph convolutional network and attention mechanism - RandGCN. This model combines piRNA sequence embedding, heterogeneous graph construction based on random walks, multi-layer graph convolution feature extraction, and multi head attention mechanism with gating units, effectively improving the accuracy and robustness of PDA prediction. The experimental results on the MNDR dataset show that RandGCN performs well in both AUC and AUPR values, demonstrating excellent predictive performance and potential applications.