Machine learning with applications最新文献

In the rapidly evolving realm of telecommunications, Machine Learning (ML) stands as a key driver for intelligent 6 G networks, leveraging diverse datasets to optimize real-time network parameters. This transition seamlessly extends from 4 G LTE and 5 G NR to 6 G, with ML insights from existing networks, specifically in predicting RRC session durations. This work introduces a novel use of weighted ensemble approach using AutoGluon library, employing multiple base models for accurate prediction of user session durations in real-world LTE and NR networks. Comparative analysis reveals superior accuracy in LTE, with 'Data Volume' as a crucial feature due to its direct impact on network load and user experience. Notably, NR sessions, marked by extended durations, reflect unique patterns attributed to Fixed Wireless Access (FWA) devices. An ablation study underscores the weighted ensemble's superior performance. This study highlights the need for techniques like data categorization to enhance prediction accuracies for evolving technologies, providing insights for enhanced adaptability in ML-based prediction models for the next network generation.

During the production of salmonids in aquaculture, it is common to observe growth-stunted individuals. The cause for the so-called “loser fish syndrome” is unclear, which needs further investigation. Here, we present and compare computer vision systems for the automatic detection and classification of loser fish in Atlantic salmon images taken in sea cages. We evaluated two end-to-end approaches (combined detection and classification) based on YoloV5 and YoloV7, and a two-stage approach based on transfer learning for detection and an ensemble of classifiers (e.g., linear perception, Adaline, C-support vector, K-nearest neighbours, and multi-layer perceptron) for classification. To our knowledge, the use of an ensemble of classifiers, considering consolidated classifiers proposed in the literature, has not been applied to this problem before. Classification entailed the assigning of every fish to a healthy and a loser class. The results of the automatic classification were compared to the reliability of human classification. The best-performing computer vision approach was based on YoloV7, which reached a precision score of 86.30%, a recall score of 71.75%, and an F1 score of 78.35%. YoloV5 presented a precision of 79.7%, while the two-stage approach reached a precision of 66.05%. Human classification had a substantial agreement strength (Fleiss’ Kappa score of 0.68), highlighting that evaluation by a human is subjective. Our proposed automatic detection and classification system will enable farmers and researchers to follow the abundance of losers throughout the production period. We provide our dataset of annotated salmon images for further research.

The application of deep neural networks towards solving problems in science and engineering has demonstrated encouraging results with the recent formulation of physics-informed neural networks (PINNs). Through the development of refined machine learning techniques, the high computational cost of obtaining numerical solutions for partial differential equations governing complicated physical systems can be mitigated. However, solutions are not guaranteed to be unique, and are subject to uncertainty caused by the choice of network model parameters. For critical systems with significant consequences for errors, assessing and quantifying this model uncertainty is essential. In this paper, an application of PINN for laser bio-effects with limited training data is provided for uncertainty quantification analysis. Additionally, an efficacy study is performed to investigate the impact of the relative weights of the loss components of the PINN and how the uncertainty in the predictions depends on these weights. Network ensembles are constructed to empirically investigate the diversity of solutions across an extensive sweep of hyper-parameters to determine the model that consistently reproduces a high-fidelity numerical simulation.

Surging in vehicle turbochargers is an important phenomenon that can damage the compressor and its peripheral equipment due to pressure fluctuations and vibration, so it is essential to understand the operating points where surging occurs. In this paper, we constructed a Neural Network (NN) that can predict these operating points, using as explanatory variables the geometry parameters of the vehicle turbocharger and one-dimensional predictions of the flow rates at surge. Our contribution is the use of machine learning to enable fast and low-cost prediction of surge points, which is usually only available through experiments or calculation-intensive Computational Fluid Dynamics (CFD). Evaluations conducted on the test data revealed that prediction accuracy was poor for some turbocharger geometries and operating conditions, and that this was associated with the relatively small data quantity included in the training data. Expanding the appropriate data offers some prospect of improving prediction accuracy.

In remote sensing, multiple input bands are derived from various sensors covering different regions of the electromagnetic spectrum. Each spectral band plays a unique role in land use/land cover characterization. For example, while integrating multiple sensors for predicting aboveground biomass (AGB) is important for achieving high accuracy, reducing the dataset size by eliminating redundant and irrelevant spectral features is essential for enhancing the performance of machine learning algorithms. This accelerates the learning process, thereby developing simpler and more efficient models. Our results indicate that compared individual sensor datasets, the random forest (RF) classification approach using recursive feature elimination (RFE) increased the accuracy based on F score by 82.86 % and 26.19 respectively. The mutual information regression (MIR) method shows a slight increase in accuracy when considering individual sensor datasets, but its accuracy decreases when all features are taken into account for all models. Overall, the combination of features from the Landsat 8, ALOS PALSAR backscatter, and elevation data selected based on RFE provided the best AGB estimation for the RF and XGBoost models. In contrast to the k-nearest neighbors (KNN) and support vector machines (SVM), no significant improvement in AGB estimation was detected even when RFE and MIR were used. The effect of parameter optimization was found to be more significant for RF than for all the other methods. The AGB maps show patterns of AGB estimates consistent with those of the reference dataset. This study shows how prediction errors can be minimized based on feature selection using different ML classifiers.

In the field of health informatics, extensive research has been conducted to predict diseases and extract valuable insights from patient data. However, a significant gap exists in addressing privacy concerns associated with data collection. Therefore, there is an urgent need to develop a machine-learning authentication model to secure the patients’ data seamlessly and continuously, as well as to find potential explanations when the model may fail. To address this challenge, we propose a unique approach to secure patients’ data using novel eigenheart features calculated from coarse-grained heart rate data. Various statistical and visualization techniques are utilized to explain the potential vulnerabilities of the model. Though it is feasible to develop continuous user authentication models from readily available heart rate data with reasonable performance, they are affected by factors such as age and Body Mass Index (BMI). These factors will be crucial for developing a more robust authentication model in the future.

Large language models (LLMs) have significantly advanced natural language processing, but their progress has yet to be equal across languages. While most LLMs are trained in high-resource languages like English, multilingual models generally underperform monolingual ones. Additionally, aspects of their multilingual foundation sometimes restrict the byproducts they produce, like computational demands and licensing regimes. In this study, we document the development of open-foundation models tailored for use in low-resource settings, their limitations, and their benefits. This is the TeenyTinyLlama pair: two compact models for Brazilian Portuguese text generation. We release them under the permissive Apache 2.0 license on GitHub and Hugging Face for community use and further development.

Large language models like ChatGPT can learn in-context (ICL) from examples. Studies showed that, due to ICL, ChatGPT achieves impressive performance in various natural language processing tasks. However, to the best of our knowledge, this is the first study that assesses ChatGPT's effectiveness in annotating a dataset for training instructor models in intelligent tutoring systems (ITSs). The task of an ITS instructor model is to automatically provide effective tutoring instruction given a student's state, mimicking human instructors. These models are typically implemented as hardcoded rules, requiring expertise, and limiting their ability to generalize and personalize instructions. These problems could be mitigated by utilizing machine learning (ML). However, developing ML models requires a large dataset of student states annotated by corresponding tutoring instructions. Using human experts to annotate such a dataset is expensive, time-consuming, and requires pedagogical expertise. Thus, this study explores ChatGPT's potential to act as a pedagogy expert annotator. Using prompt engineering, we created a list of instructions a tutor could recommend to a student. We manually filtered this list and instructed ChatGPT to select the appropriate instruction from the list for the given student's state. We manually analyzed ChatGPT's responses that could be considered incorrectly annotated. Our results indicate that using ChatGPT as an annotator is an effective alternative to human experts. The contributions of our work are (1) a novel dataset annotation methodology for the ITS, (2) a publicly available dataset of student states annotated with tutoring instructions, and (3) a list of possible tutoring instructions.

Maize is an important crop cultivated in Sub-Saharan Africa, essential for food security. However, its cultivation faces significant challenges due to debilitating diseases such as Maize Lethal Necrosis (MLN) and Maize Streak Virus (MSV), which can lead to severe yield losses. Traditional plant disease diagnosis methods are often time-consuming and prone to errors, necessitating more efficient approaches. This study explores the application of deep learning, specifically Convolutional Neural Networks (CNNs), in the automatic detection and classification of maize diseases. We investigate six architectures: Basic CNN, EfficientNet V2 B0 and B1, LeNet-5, VGG-16, and ResNet50, using a dataset of 15344 images comprising MSV, MLN, and healthy maize leaves. Additionally, We performed hyperparameter tuning to improve the performance of the models and Gradient-weighted Class Activation Mapping (Grad-CAM) for model interpretability. Our results show that the EfficientNet V2 B0 model demonstrated an accuracy of 99.99% in distinguishing between healthy and disease-infected plants. The results of this study contribute to the advancement of AI applications in agriculture, particularly in diagnosing maize diseases within Sub-Saharan Africa.

Breast cancer is a condition where the irregular growth of breast cells occurs uncontrollably, leading to the formation of tumors. It poses a significant threat to women’s lives globally, emphasizing the need for enhanced methods of detecting and categorizing the disease. In this work, we propose an Ensemble Deep Convolutional Neural Network (EDCNN) model that exhibits superior accuracy compared to several transfer learning models and the Vision Transformer model. Our EDCNN model integrates the strengths of the MobileNet and Xception models to improve its performance in breast cancer detection and classification. We employ various preprocessing techniques, including image resizing, data normalization, and data augmentation, to prepare the data for analysis. By following these measures, the formatting is optimized, and the model’s capacity to make generalizations is improved. We trained and evaluated our proposed EDCNN model using ultrasound images, a widely available modality for breast cancer imaging. The outcomes of our experiments illustrate that the EDCNN model attains an exceptional accuracy of 87.82% on Dataset 1 and 85.69% on Dataset 2, surpassing the performance of several well-known transfer learning models and the Vision Transformer model. Furthermore, an AUC value of 0.91 on Dataset 1 highlights the robustness and effectiveness of our proposed model. Moreover, we highlight the incorporation of the Grad-CAM Explainable Artificial Intelligence (XAI) technique to improve the interpretability and transparency of our proposed model. Additionally, we performed image segmentation using the U-Net segmentation technique on the input ultrasound images. This segmentation process allowed for the identification and isolation of specific regions of interest, facilitating a more comprehensive analysis of breast cancer characteristics. In conclusion, the study presents a creative approach to detecting and categorizing breast cancer, demonstrating the superior performance of the EDCNN model compared to well-established transfer learning models. Through advanced deep learning techniques and image segmentation, this study contributes to improving diagnosis and treatment outcomes in breast cancer.