Overfitting Issues Research Articles

Enhancing image-based diagnosis of gastrointestinal tract diseases through deep learning with EfficientNet and advanced data augmentation techniques.

The early detection and diagnosis of gastrointestinal tract diseases, such as ulcerative colitis, polyps, and esophagitis, are crucial for timely treatment. Traditional imaging techniques often rely on manual interpretation, which is subject to variability and may lack precision. Current methodologies leverage conventional deep learning models that, while effective to an extent, often suffer from overfitting and generalization issues on medical image datasets due to the intricate and subtle variations in disease manifestations. These models typically do not fully utilize the potential of transfer learning or advanced data augmentation, leading to less-than-optimal performance, especially in diverse real-world scenarios where data variability is high. This study introduces a robust model using the EfficientNetB5 architecture combined with a sophisticated data augmentation strategy. The model is tailored for the high variability and intricate details present in gastrointestinal tract disease images. By integrating transfer learning with maximal pooling and extensive regularization, the model aims to enhance diagnostic accuracy and reduce overfitting. The proposed model achieved a test accuracy of 98.89%, surpassing traditional methods by incorporating advanced regularization and augmentation techniques. The application of horizontal flipping and dynamic scaling during training significantly improved the model's ability to generalize, evidenced by a low-test loss of 0.230 and high precision metrics across all classes. The proposed deep learning framework demonstrates superior performance in the automated classification of gastrointestinal diseases from image data. By addressing key limitations of existing models through innovative techniques, this study contributes to the enhancement of diagnostic processes in medical imaging, potentially leading to more accurate and timely disease interventions.

Detection of Subarachnoid Hemorrhage Using CNN with Dynamic Factor and Wandering Strategy-Based Feature Selection.

Objectives: Subarachnoid Hemorrhage (SAH) is a serious neurological emergency case with a higher mortality rate. An automatic SAH detection is needed to expedite and improve identification, aiding timely and efficient treatment pathways. The existence of noisy and dissimilar anatomical structures in NCCT images, limited availability of labeled SAH data, and ineffective training causes the issues of irrelevant features, overfitting, and vanishing gradient issues that make SAH detection a challenging task. Methods: In this work, the water waves dynamic factor and wandering strategy-based Sand Cat Swarm Optimization, namely DWSCSO, are proposed to ensure optimum feature selection while a Parametric Rectified Linear Unit with a Stacked Convolutional Neural Network, referred to as PRSCNN, is developed for classifying grades of SAH. The DWSCSO and PRSCNN surpass current practices in SAH detection by improving feature selection and classification accuracy. DWSCSO is proposed to ensure optimum feature selection, avoiding local optima issues with higher exploration capacity and avoiding the issue of overfitting in classification. Firstly, in this work, a modified region-growing method was employed on the patient Non-Contrast Computed Tomography (NCCT) images to segment the regions affected by SAH. From the segmented regions, the wide range of patterns and irregularities, fine-grained textures and details, and complex and abstract features were extracted from pre-trained models like GoogleNet, Visual Geometry Group (VGG)-16, and ResNet50. Next, the PRSCNN was developed for classifying grades of SAH which helped to avoid the vanishing gradient issue. Results: The DWSCSO-PRSCNN obtained a maximum accuracy of 99.48%, which is significant compared with other models. The DWSCSO-PRSCNN provides an improved accuracy of 99.62% in CT dataset compared with the DL-ICH and GoogLeNet + (GLCM and LBP), ResNet-50 + (GLCM and LBP), and AlexNet + (GLCM and LBP), which confirms that DWSCSO-PRSCNN effectively reduces false positives and false negatives. Conclusions: the complexity of DWSCSO-PRSCNN was acceptable in this research, for while simpler approaches appeared preferable, they failed to address problems like overfitting and vanishing gradients. Accordingly, the DWSCSO for optimized feature selection and PRSCNN for robust classification were essential for handling these challenges and enhancing the detection in different clinical settings.

Enhancing Efficacy in Breast Cancer Screening with Nesterov Momentum Optimization Techniques

In the contemporary landscape of healthcare, machine learning models are pivotal in facilitating precise predictions, particularly in the nuanced diagnosis of complex ailments such as breast cancer. Traditional diagnostic methodologies grapple with inherent challenges, including excessive complexity, elevated costs, and reliance on subjective interpretation, which frequently culminate in inaccuracies. The urgency of early detection cannot be overstated, as it markedly broadens treatment modalities and significantly enhances survival rates. This paper delineates an innovative optimization framework designed to augment diagnostic accuracy by amalgamating momentum-based optimization techniques within a neural network paradigm. Conventional machine learning approaches are often encumbered by issues of overfitting, data imbalance, and the inadequacy of capturing intricate patterns in high-dimensional datasets. To counter these limitations, we propose a sophisticated framework that integrates an adaptive threshold mechanism across an array of gradient-based optimizers, including SGD, RMSprop, adam, adagrad, adamax, adadelta, nadam and Nesterov momentum. This novel approach effectively mitigates oscillatory behavior, refines parameter updates, and accelerates convergence. A salient feature of our methodology is the incorporation of a momentum threshold for early stopping, which ceases training upon the stabilization of momentum below a pre-defined threshold, thereby pre-emptively preventing overfitting. Leveraging the Wisconsin Breast Cancer Dataset, our model achieved a remarkable 99.72% accuracy and 100% sensitivity, significantly curtailing misclassification rates compared to traditional methodologies. This framework stands as a robust solution for early breast cancer diagnosis, thereby enhancing clinical decision making and improving patient outcomes.

Road Pothole Detection Model Based on Local Attention Resnet18-CNN-LSTM

Abstract. In response to the low detection accuracy and slow speed of existing road pothole detection methods, a road pothole classification detection model based on local attention Resnet18-CNN-LSTM (Long Short-Term Memory network) is proposed. On the basis of Resnet18, a local attention mechanism and a CNN-LSTM combined model are added to propose a road pothole detection model based on local attention Resnet18-CNN-LSTM. The local attention mechanism is used to accurately extract specific target feature values, CNN is used to extract the spatial features of the input data, and LSTM enhances the detection model's extraction of sequential features and performs classification, thereby improving the accuracy of the road and pothole model. Experimental results show that on the training set, the accuracy of the local attention mechanism-based ResNet18-CNN-LSTM model reached 99.2188%, which is an increase of 0.7813% and 2.3438% compared to the ResNet34-CNN-LSTM and ResNet50-CNN-LSTM models under the same conditions, respectively. On the test set, the model's accuracy was 93.4437%, an increase of 0.5437% and 1.9867% compared to the ResNet34-CNN-LSTM and ResNet50-CNN-LSTM models, respectively. After dealing with overfitting issues through early stopping, the detection accuracy of this model has significantly improved compared to the detection models based on ResNet34 and ResNet50, with an increase of 1.2% and 1.49% respectively. The model shows faster processing speed in identification time, effectively retains the correlation and sequence features of the data, overcomes the problem of gradient disappearance in deep networks, and thereby enhances the extraction capability of local target features of road pothole images. The above results indicate that the local attention mechanism-based ResNet18-CNN-LSTM model shows superior performance in road pothole detection.

Enhanced Short-Term Load Forecasting: Error-Weighted and Hybrid Model Approach

To tackle the challenges of high variability and low accuracy in short-term electricity load forecasting, this study introduces an enhanced prediction model that addresses overfitting issues by integrating an error-optimal weighting approach with an improved ensemble forecasting framework. The model employs a hybrid algorithm combining grey relational analysis and radial kernel principal component analysis to preprocess the multi-dimensional input data. It then leverages an ensemble of an optimized deep bidirectional gated recurrent unit (BiGRU), an enhanced long short-term memory (LSTM) network, and an advanced temporal convolutional neural network (TCN) to generate predictions. These predictions are refined using an error-optimal weighting scheme to yield the final forecasts. Furthermore, a Bayesian-optimized Bagging and Extreme Gradient Boosting (XGBoost) ensemble model is applied to minimize prediction errors. Comparative analysis with existing forecasting models demonstrates superior performance, with an average absolute percentage error (MAPE) of 1.05% and a coefficient of determination (R2) of 0.9878. These results not only validate the efficacy of our proposed strategy, but also highlight its potential to enhance the precision of short-term load forecasting, thereby contributing to the stability of power systems and supporting societal production needs.

Automatic damage detection and segmentation using deep learning algorithms in reinforced concrete structure inspections

AbstractTraditional methods of detecting concrete damage, such as manual inspection, are typically slow, labor‐intensive, and subjective. Integrating deep learning algorithms has automated this process, representing a significant advance in building damage inspection. Semantic segmentation, a technique within deep learning, is increasingly recognized for its capability to identify the location and shape of concrete damage accurately. Unlike basic deep learning approaches like classification and object detection, semantic segmentation not only recognizes damage but also delineates its boundaries, and has great potential in facilitating dimension measurement. However, multi‐level, high‐precision detection for various structural damage assessments remains an area requiring further research. This study created a database of reinforced concrete surfaces with pixel‐level, multi‐category semantic segmentation annotations for various levels of component damage, including cracks and spalling of concrete, exposure, buckling, fracture of reinforcing bars, etc. The performance of advanced deep learning segmentation algorithms, including U‐Net, DeepLab, K‐Net, and FastSCNN, was consequently trained and evaluated in detecting various types of concrete damage. All models show good accuracy of more than 98%, but lower F1 scores around 70%. U‐Net and K‐Net demonstrate relatively stable performance, indicating a certain degree of consistency and a high‐performance peak. In contrast, DeepLabv3's F1 score fluctuates significantly, suggesting the model may suffer from overfitting or other stability issues during training, resulting in a relatively low average performance. FastSCNN exhibits the highest potential, achieving the highest F1 score. In addition, this study also tested the performance of each model on new damage images and explored the impact of the dataset proportion ratio (training set vs. validation set), contributing to providing insights into each algorithm's suitability for various inspection scenarios. Finally, perspectives on current challenges and future directions in the field have also been given.

An Unbiased Risk Estimator for Partial Label Learning with Augmented Classes

Partial Label Learning (PLL) is a typical weakly supervised learning task, which assumes each training instance is annotated with a set of candidate labels containing the ground-truth label. Recent PLL methods adopt identification-based disambiguation to alleviate the influence of false positive labels and achieve promising performance. However, they require all classes in the test set to have appeared in the training set, ignoring the fact that new classes will keep emerging in real applications. To address this issue, in this paper, we focus on the problem of Partial Label Learning with Augmented Class (PLLAC), where one or more augmented classes are not visible in the training stage but appear in the inference stage. Specifically, we propose an unbiased risk estimator with theoretical guarantees for PLLAC, which estimates the distribution of augmented classes by differentiating the distribution of known classes from unlabeled data and can be equipped with arbitrary PLL loss functions. Besides, we provide a theoretical analysis of the estimation error bound of the estimator, which guarantees the convergence of the empirical risk minimizer to the true risk minimizer as the number of training data tends to infinity. Furthermore, we add a risk-penalty regularization term in the optimization objective to alleviate the influence of the over-fitting issue caused by negative empirical risk. Extensive experiments on benchmark, UCI and real-world datasets demonstrate the effectiveness of the proposed approach.

Utilizing Bayesian generalization network for reliable fault diagnosis of machinery with limited data

To address the issues of overfitting, domain generalization challenges, and lack of credibility brought by limited data samples in mechanical fault diagnosis in practical engineering, this paper proposes a reliable Bayesian generalization network (BGNet). A Bayesian convolutional layer is constructed based on variational inference, treating all parameters in the convolutional layer as random variables. This approach makes a single model function similar to an ensemble of an infinite number of models, and thus enhancing the model's capability of overfitting resistance and domain generalization. The parameters of the variational distribution are updated to approximate the posterior distribution by local reparametrization and Monte Carlo sampling to optimize the evidence lower bound (ELBO) loss. Confidence information is extracted from the model results and, uncertainty estimation and decomposition schemes are designed to provide interpretability. The proposed method is applied to analyze the experimental data of bearing and gearbox faults. The results show that in a multi-source domain scenario with limited samples, the proposed method demonstrates high diagnostic accuracy, effectively describes the relationship between domain variability and uncertainty, and significantly outperforms several benchmark and state-of-the-art models.

Exploratory parallel hybrid sampling framework for imbalanced data classification

Current engineering application scenarios often face the challenge of imbalanced data, hybrid sampling is an effective method to deal with the imbalanced data classification issue, which can avoid the issues of overfitting and mistakenly deleting useful majority samples when using oversampling approach and undersampling approach alone. However, at present most of the hybrid sampling approaches are implemented serially, and the implementation of oversampling and undersampling approaches alone will cause mutual interference and influence between them. This study proposes a parallel hybrid sampling framework based on the idea of parallel engineering and theoretically analyzes its superiority. The experimental results show that when applied to five classification algorithms with three performance evaluation metrics,the proposed framework outperforms the two mainstream hybrid sampling frameworks. Moreover, the proposed framework can effectively reduce the time consumption of hybrid sampling process.

Artificial intelligent-driven decision-making for automating root fracture detection in periapical radiographs

BackgroundThe detection and early diagnosis of root fractures can be challenging; this difficulty applies particularly to newly qualified dentists. Aside from clinical examination, diagnosis often requires radiographic assessment. Nonetheless, human fallibility can introduce errors due to a lack of experience.AimThe proposed system aimed to assist in detecting root fractures through the integration of artificial intelligence techniques into the diagnosis process as a step for automating dental diagnosis and decision-making processes.Materials and methodA total of 400 radiographic images of fractured and unfractured teeth were obtained for the present research. Data handling techniques were implemented to balance the distribution of the samples. The AI-based system used the voting technique for five different pretrained models namely, VGG16, VGG19, ResNet50. DenseNet121, and DenseNet169 to perform the analysis. The parameters used for the analysis of the models are loss and accuracy curves.ResultsVGG16 exhibited notable success with low training and validation losses (0.09% and 0.18%, respectively), high specificity, sensitivity, and positive predictive value (PPV). VGG19 showed potential overfitting concerns, while ResNet50 displayed progress in minimizing loss but exhibited bias toward unfractured cases. DenseNet121 effectively addressed overfitting and noise issues, achieving balanced metrics and impressive PPVs for both fractured and unfractured cases (0.933 and 0.898 respectively). With increased depth, DenseNet169 demonstrated enhanced generalization capability.ConclusionThe proposed AI- based system demonstrated high precision and sensitivity for detecting root fractures in endodontically treated teeth by utilizing the voting method.

HcGAN: Harmonic conditional generative adversarial network for efficiently generating high-quality IHC images from H&E

Generating high quality histopathology images like immunohistochemistry (IHC) stained images is essential for precise diagnosis and the advancement of computer-aided diagnostic (CAD) systems. Producing IHC images in laboratory is quite expensive and time consuming. Recently, some attempts have been made based on artificial intelligence techniques (particularly, deep learning) to generate IHC images. Existing IHC stained image generation methods, still have a limited performance due to the complex structures and variations in cells shapes, potential nonspecific staining and variable antibody sensitivity. This paper proposes a novel technique known as harmonic conditional generative adversarial network (HcGAN) for generating high quality IHC-stained images. To generate the IHC images, the HcGAN model is fed with the widely available hematoxylin and eosin (H&E) images that contain cellular and morphological underlying structures of diverse cancer tissues. Such approach helps generate high quality IHC images mimicking the real ones that highlight the positive cells. The proposed HcGAN model is based generative adversarial learning with generator and discriminator networks. In HcGAN, harmonic convolution based on discrete cosine transform filter banks is employed in the generator and discriminator networks instead of the standard convolution in order to improve visual quality of the generated images and address the issue of overfitting. Our qualitative and quantitative results demonstrate that the proposed HcGAN achieved the highest performance over state-of-the-art methods using two publicly available datasets.

Text-free diffusion inpainting using reference images for enhanced visual fidelity

This paper presents a novel approach to subject-driven image generation that addresses the limitations of traditional text-to-image diffusion models. Our method generates images using reference images without relying on language-based prompts. We introduce a visual detail preserving module that captures intricate details and textures, addressing overfitting issues associated with limited training samples. The model's performance is further enhanced through a modified classifier-free guidance technique and feature concatenation, enabling the natural positioning and harmonization of subjects within diverse scenes. Quantitative assessments using CLIP, DINO and Quality scores (QS), along with a user study, demonstrate the superior quality of our generated images. Our work highlights the potential of pre-trained models and visual patch embeddings in subject-driven editing, balancing diversity and fidelity in image generation tasks. Our implementation is available at https://github.com/8eomio/Subject-Inpainting.

Comparison of RNN-LSTM, TFDF and stacking model approach for weather forecasting in Bangladesh using historical data from 1963 to 2022

Forecasting the weather in an area characterized by erratic weather patterns and unpredictable climate change is a challenging endeavour. The weather is classified as a non-linear system since it is influenced by various factors that contribute to climate change, such as humidity, average temperature, sea level pressure, and rainfall. A reliable forecasting system is crucial in several industries, including transportation, agriculture, tourism, & development. This study showcases the effectiveness of data mining, meteorological analysis, and machine learning techniques such as RNN-LSTM, TensorFlow Decision Forest (TFDF), and model stacking (including ElasticNet, GradientBoost, KRR, and Lasso) in improving the precision and dependability of weather forecasting. The stacking model strategy entails aggregating multiple base models into a meta-model to address issues of overfitting and underfitting, hence improving the accuracy of the prediction model. To carry out the study, a comprehensive 60-year meteorological record from Bangladesh was gathered, encompassing data on rainfall, humidity, average temperature, and sea level pressure. The results of this study suggest that the stacking average model outperforms the TFDF and RNN-LSTM models in predicting average temperature. The stacking average model achieves an RMSLE of 1.3002, which is a 10.906% improvement compared to the TFDF model. It is worth noting that the TFDF model had previously outperformed the RNN-LSTM model. The performance of the individual stacking model is not as impressive as that of the average model, with the validation results being better in TFDF.

A novel integrated logistic regression model enhanced with recursive feature elimination and explainable artificial intelligence for dementia prediction

Dementia is a major global health issue that significantly impacts millions of individuals, families, and societies worldwide, creating a substantial burden on healthcare systems. This study introduces a novel approach for predicting dementia by employing the Logistic Regression (LR) model, enhanced with Recursive Feature Elimination (RFE), applied to a unique dataset comprising 1000 patients, with 49.60% male and 50.40% female. The LR model, recognized for its simplicity and effectiveness in binary classification tasks, is optimized through RFE, a technique that iteratively eliminates less significant features to improve model performance. The model’s effectiveness was assessed using comprehensive metrics, including accuracy, precision, recall, F1-score, Matthews Correlation Coefficient (MCC), and Kappa score. Furthermore, SHapley Additive exPlanations (SHAP) values were employed to increase the interpretability of the model, providing insights into the most influential features for dementia prediction. To address the issue of overfitting, a standardization technique was implemented, which enhanced the model’s predictive performance. The findings of this study hold potential implications for early dementia detection, informing intervention strategies, and optimizing healthcare resource allocation.

Unpacking the Gap Box Against Data-Free Knowledge Distillation.

Data-free knowledge distillation (DFKD) improves the student model (S) by mimicking the class probability from a pre-trained teacher model (T) without training data. Under such setting, an ideal scenario is that T can help generate "good" samples from a generator (G) to maximally benefit S. However, existing arts suffer from the non-ideal generated samples under the disturbance of the gap (i.e., either too large or small) between the class probabilities of T and S; for example, the generated samples with too large gap may exhibit excessive information for S, while too small gap leads to the limited knowledge in the samples, resulting into the poor generalization. Meanwhile, they fail to judge the "goodness" of the generated samples for S since the fixed T is not necessarily ideal. In this paper, we aim to answer what is inside the gap box; together with how to yield "good" generated samples for DFKD? To this end, we propose a Gap-Sensitive Sample Generation (GapSSG) approach, by revisiting the empirical distilled risk from a data-free perspective, which confirms the existence of an ideal teacher (T *), while theoretically implying: (1) the gap disturbance originates from the mismatch between T and T *, hence the class probabilities of T enable the approximation to those of T *; and (2) "good" samples should maximally benefit S via T's class probabilities, owing to unknown T *. To this end, we unpack the gap box between T and S as two findings: inherent gap to perceive T and T *; derived gap to monitor S and T *. Benefiting from the derived gap that focuses on the adaptability of generated sample to S, we attempt to track student's training route (a series of training epochs) to capture the category distribution of S; upon which, a regulatory factor is further devised to approximate T * over inherent gap, so as to generate "good" samples to S. Furthermore, during the distillation process, a sample-balanced strategy comes up to tackle the overfitting and missing knowledge issues between the generated partial and critical samples by training G. The theoretical and empirical studies verify the advantages of GapSSG over the state-of-the-arts.

EEG classification with limited data: A deep clustering approach

In this paper, we propose a novel training strategy designed to prevent deep neural networks from overfitting when trained on a very limited number of samples. This approach enables the use of very deep networks in applications like EEG classification, where data collection is challenging, leading to significant overfitting issues. To address this issue, we introduce a classification network that leverages a deep clustering operation in its latent space. The network utilizes a set of dictionary elements as auxiliary inputs and learns to guide each input sample toward one of the clusters, each derived from a dictionary element. This process facilitates the grouping of samples from the same class within the same region of the latent space, thereby reducing the risk of complex class boundaries and overfitting. Additionally, we introduce two probabilistic models within our clustering approach to mitigate outlier issues and address augmentation challenges commonly encountered with EEG signals. Experimental evaluations on two widely recognized databases, EEG-ImageNet and BCIIV2a, demonstrate the effectiveness and superiority of our method compared to state-of-the-art approaches, achieving 97.9% accuracy on EEG-ImageNet and 84.1% accuracy on BCIIV2a.

Quantitative modeling and uncertainty estimation for small-sample LIBS using Gaussian negative log-likelihood and monte carlo dropout methods

Laser-induced breakdown spectroscopy (LIBS) has emerged as a powerful analytical technique widely applied in materials science, environmental monitoring, and other fields. In traditional quantitative analysis, models are established for the sample-to-element content based on acquired characteristic spectra combined with specific modeling methods. Small-sample LIBS poses significant challenges to researchers due to its high dimensionality and limited sample size, making it difficult for traditional modeling methods to establish effectively generalized and robust models. Addressing the issues of overfitting during training and the uncertainty of LIBS data, we developed an artificial neural network (ANN) model. Unlike traditional ANN which uses mean squared error functions as training losses, Gaussian negative log-likelihood (GLL) functions were used to train the model. This approach allows for simultaneous consideration of the quantitative prediction of sample element concentrations and the uncertainty of the quantification model prediction. Moreover, the Monte Carlo Dropout (MCDropout) method was applied to reduce prediction bias, resulting in a more robust, generalizable, and uncertainty-aware quantitative model. Compared to traditional quantitative methods, the method proposed in this paper achieved quantitative accuracies of 0.9877, 0.9939, 0.9876, and 0.9899 for four elements (Mn, Mo, Cr, Cu) in stainless steel, respectively. The uncertainty output effectively informs data comprehension and spectral analysis. To the best of our knowledge, this study is the first to propose utilizing model-quantified spectral uncertainty in LIBS quantitative tasks. This innovative study not only provides solutions for small-sample LIBS learning but also enhances the stability and accuracy of quantitative modeling by addressing the impact of uncertainty on training and reducing uncertainty. It is crucial for overcoming the limitations of traditional methods in small-sample LIBS quantification and achieving reliable real-time, high-precision LIBS detection, taking LIBS technology a step further towards practical reliability and precision.

Vessel trajectory classification via transfer learning with Deep Convolutional Neural Networks.

The classification of vessel trajectories using Automatic Identification System (AIS) data is crucial for ensuring maritime safety and the efficient navigation of ships. The advent of deep learning has brought about more effective classification methods, utilizing Convolutional Neural Networks (CNN). However, existing CNN-based approaches primarily focus on either sailing or loitering movement patterns and struggle to capture valuable features and subtle differences between these patterns from input images. In response to these limitations, we firstly introduce a novel framework, Dense121-VMC, based on Deep Convolutional Neural Networks (DCNN) with transfer learning for simultaneous extraction and classification of both sailing and loitering trajectories. Our approach efficiently performs in extracting significant features from input images and in identifying subtle differences in each vessel's trajectory. Additionally, transfer learning effectively reduces data requirements and addresses the issue of overfitting. Through extended experiments, we demonstrate the novelty of proposed Dense121-VMC framework, achieving notable contributions for vessel trajectory classification.

RDMS: Reverse distillation with multiple students of different scales for anomaly detection

AbstractUnsupervised anomaly detection, often approached as a one‐class classification problem, is a critical task in computer vision. Knowledge distillation has emerged as a promising technique for enhancing anomaly detection accuracy, especially with the advent of reverse distillation networks that employ encoder–decoder architectures. This study introduces a novel reverse knowledge distillation framework known as RDMS, which incorporates a pretrained teacher encoding module, a multi‐level feature fusion connection module, and a student decoding module consisting of three independent decoders. RDMS is designed to distill distinct features from the teacher encoder, mitigating overfitting issues associated with similar or identical teacher–student structures. The model achieves an average of 99.3 image‐level AUROC and 98.34 pixel‐level AUROC on the MVTec‐AD dataset and demonstrates state‐of‐the‐art performance on the more challenging BTAD dataset. The RDMS model's high accuracy in anomaly detection and localization underscores the potential of multi‐student reverse distillation to advance unsupervised anomaly detection capabilities. The source code is available at https://github.com/zihengchen777/RDMS

Optimisation of the adaptive neuro-fuzzy inference system for adjusting low-cost sensors PM concentrations

Driven by the urgent necessity for accurate environmental data in urban settings, this research leverages the Adaptive Neuro-Fuzzy Inference System (ANFIS) as a machine learning-based approach to refine SPS30 low-cost sensor data influenced by hygroscopicity in Turin, Italy. Employing ANFIS offers several advantages: it enhances clarity regarding the correspondence between output and input values and rules, improves system interpretability, and facilitates the representation of linguistic variables and rules, thereby encouraging domain experts' involvement in enhancing the system's performance as needed. This paper illustrates the utility of ANFIS in adjusting the detected particulate matter (PM) concentration and compares its effectiveness with other established machine-learning techniques, including linear regression, decision trees, random forest, SVR and a multilayer perceptron (MLP). These methods are chosen as benchmarks owing to their established effectiveness in calibration procedures.We propose certain preprocessing steps for detecting and rectifying anomalies, alongside introducing two distinct data-splitting methodologies. Additionally, a discussion about feature selection is presented to elucidate the impact of specific features on performance enhancement. The efficacy of ANFIS in refining PM data is demonstrated through a comparative assessment, where it outperforms all the established machine-learning techniques. Notably, incorporating only PM2.5, relative humidity and temperature as features yields optimal performance while mitigating overfitting issues. The paper also explores various ANFIS configurations, including two distinct optimization algorithms, and investigates the impact of the number and type of membership functions on the fuzzy system's performance. Our study highlights the potential of the Adaptive Neuro-Fuzzy Inference System as a versatile and effective tool for addressing real-world challenges in environmental sensing.