Amount Of Training Data Research Articles

Optimized decomposition and identification method for multiple power quality disturbances

AbstractThe complexity of power quality (PQ) concerns is intensifying in tandem with the proliferation of inverter‐based renewable energy systems. The integration of power electronic devices within the distribution network exacerbates the complexity and introduces greater temporal variability to signal components. This paper introduces an advanced, online optimization technique for the decomposition and identification of PQ disturbances (PQDs). Initially, an improved variational mode decomposition (IVMD) method is presented, leveraging an energy ratio criterion for precise decomposition of concurrent PQDs. Subsequently, utilizing the characteristic attributes derived from IVMD, an optimized support vector machine (OSVM) algorithm is developed through the synthesis of diverse kernel functions. The OSVM strategically employs distinct kernel functions to augment the discriminability of the feature set. The synergy of IVMD and OSVM enables the detection of multiple PQDs, remarkably even with a minimal amount of training data. A series of experiments have been conducted to validate the effectiveness of the proposed methodology. The results corroborate that the formulated framework exhibits robust learning capabilities and a high degree of resistance to noise interference. Moreover, the hardware platform experiments prove that the proposed method has a satisfactory real‐time performance for its practicability.

Lensless Imaging Based on Dual‐Input Physics‐Driven Neural Network

Lensless imaging, as a novel computational imaging technique, has attracted great attention due to its simplicity, compactness, and flexibility. This technique analyzes and processes the diffraction of an object to obtain complex amplitude information. However, traditional algorithms such as Gerchberg‐Saxton (G–S) algorithm tend to exhibit significant errors in complex amplitude retrieval, particularly for edge information. Additional constraints have to be incorporated on top of amplitude constraints to enhance the accuracy. Recently, deep learning has shown promising results in optical imaging. However, it requires a large amount of training data. To address these issues, a novel approach called dual‐input physics‐driven network (DPNN) is proposed for lensless imaging. DPNN utilizes two diffractions recorded at different distances as inputs and uses an unsupervised approach that combines physical imaging model to reconstruct object information. DPNN adopts a U‐Net 3+ architecture with a loss function of mean absolute error (MAE) to better capture diffraction features. DPNN achieves highly accurate reconstruction without requiring extensive data and being immune to background noise. Based on different diffraction intervals, noise levels, and imaging models, DPNN exhibits superior capabilities in peak signal‐to‐noise ratio and structural similarity compared with conventional methods, effectively achieving accurate phase or amplitude information reconstruction.

Patching-based deep-learning model for the inpainting of Bragg coherent diffraction patterns affected by detector gaps.

A deep-learning algorithm is proposed for the inpainting of Bragg coherent diffraction imaging (BCDI) patterns affected by detector gaps. These regions of missing intensity can compromise the accuracy of reconstruction algorithms, inducing artefacts in the final result. It is thus desirable to restore the intensity in these regions in order to ensure more reliable reconstructions. The key aspect of the method lies in the choice of training the neural network with cropped sections of diffraction data and subsequently patching the predictions generated by the model along the gap, thus completing the full diffraction peak. This approach enables access to a greater amount of experimental data for training and offers the ability to average overlapping sections during patching. As a result, it produces robust and dependable predictions for experimental data arrays of any size. It is shown that the method is able to remove gap-induced artefacts on the reconstructed objects for both simulated and experimental data, which becomes essential in the case of high-resolution BCDI experiments.

StereoThermoLegs: label propagation with multimodal stereo cameras for automated annotation of posterior legs during running at different velocities

In sports science, thermal imaging is applied to investigate various questions related to exercise-induced stress response, muscle fatigue, anomalies, and diseases. Infrared thermography monitors thermal radiation from the skin’s surface over time. For further analysis, regions of interest are extracted and statistically analyzed. Although computer vision algorithms have grown in recent years due to data-driven approaches, this is not the case for detailed segmentation in thermal images. In a supervised manner, machine learning optimizations require a large amount of training data with input and ground truth output data. Unfortunately, obtaining annotated data are a costly problem that increases with the complexity of the task. For semantic segmentation, pixel-wise label masks must be created by experts. Few datasets meet the needs of sports scientists and physicians to perform advanced applications of thermal computer vision during physical activity and generate new insights in their fields. In this paper, a new method is introduced to transfer segmentation masks from the vision domain to the thermal domain with a stereo-calibrated time-of-flight camera and high-resolution mid-wave infrared camera. A post-processing procedure is then utilized to obtain dense pixel masks for the posterior legs during walking and running on a treadmill. The developed StereoThermoLegs dataset is based on 14 participants and includes 11 subjects for training with 12,826 thermograms and the remaining three individuals for testing with 3433 images. A deep neural network was trained with the DeepLabv3+ architecture, the AdaBelief optimizer, and Dice loss as a benchmark. After 29 epochs, the test set achieved an average intersection over union of 0.66. The analysis of the posterior leg region, specifically the left and right calf, offered the most insights, with values of 0.83 and 0.83, respectively. The first multimodal stereo dataset containing synchronized visual and thermal images of a runner’s back provides a starting point for data-driven segmentation tasks in sports science and medicine. Our technique allows for automatic production of customized datasets for deep learning, accelerating the implementation of baseline outcomes for newly identified areas of interest in thermal imaging, while bypassing the requirement for extensive manual annotation. The approach is not exclusive to stereo rig and segmentation tasks utilizing RGBD and thermal cameras, but can be applied to other imaging tasks and modalities.

Combined classification and regression inverse kinematics modeling for overall shape prediction of soft manipulators

In helping astronauts perform tasks, soft manipulators offer the advantages of softness, lightness, low cost, and interaction safety to interact with. However, modeling their inverse kinematics through traditional methods is subject to truncation errors, complex derivations and inexpressibility equations, owing to geometric singularities. Furthermore, facing the future control requirement for multiple major feature points on manipulators, most neural network approaches that solve the drive space only by the inverse solution of the end task space, cannot obtain the inverse kinematics solutions of all feature points, due to their including redundancy (the dimension n of the drive space of feature points is larger than that of the end task space). Therefore, this paper proposes a method combining classification and regression based on multi-layer perceptron for inverse kinematic modeling, namely the IK-MLP-CR method, with the bending states classified and the whole manipulator shape predicted, greatly reducing the amount of training data and training time. Finally, the performance evaluation of the proposed model is validated through the simulation and ground microgravity experiments. Results show that using the IK-MLP-CR method results in a successful grasping rate of over 97%, and that the trend of the experimental bending conforms to physical motions under different pressures.

Selective privacy-preserving framework for large language models fine-tuning

Fine-tuning pre-trained large language models (LLMs) helps various downstream tasks, but brings serious privacy leaks when relying on large amounts of data for training. Differentially private stochastic gradient descent (DPSGD) has been designed to introduce noise during model updates to prevent privacy leaks. Nevertheless, fine-tuning LLMs via DPSGD limits the model utility since heavy perturbations are introduced on large high-dimensional gradients. Besides, existing privacy-preserving mechanisms directly perturb all tokens of the input sentences, which are too pessimistic to achieve good model performance. Therefore, this paper researches a selective privacy-preserving framework for fine-tuning LLMs. We propose a first-of-its-kind privacy notion called selective sequence local differential privacy (S-SeqLDP), which provides guarantees of indistinguishability only for the secret part of the sequences. Furthermore, we design a novel framework called SLDP-FT that enables S-SeqLDP-compliant large language model fine-tuning by perturbing the forward-pass embeddings with selective noises. We innovatively investigate the privacy forward weight that determines the noise magnitude of achieving selective privacy protection. Extensive experiments on three tasks demonstrate that our SLDP-FT achieves better model accuracy than state-of-the-art techniques when providing the same level of privacy protection.

End-to-end pseudonymization of fine-tuned clinical BERT models

Many state-of-the-art results in natural language processing (NLP) rely on large pre-trained language models (PLMs). These models consist of large amounts of parameters that are tuned using vast amounts of training data. These factors cause the models to memorize parts of their training data, making them vulnerable to various privacy attacks. This is cause for concern, especially when these models are applied in the clinical domain, where data are very sensitive. Training data pseudonymization is a privacy-preserving technique that aims to mitigate these problems. This technique automatically identifies and replaces sensitive entities with realistic but non-sensitive surrogates. Pseudonymization has yielded promising results in previous studies. However, no previous study has applied pseudonymization to both the pre-training data of PLMs and the fine-tuning data used to solve clinical NLP tasks. This study evaluates the effects on the predictive performance of end-to-end pseudonymization of Swedish clinical BERT models fine-tuned for five clinical NLP tasks. A large number of statistical tests are performed, revealing minimal harm to performance when using pseudonymized fine-tuning data. The results also find no deterioration from end-to-end pseudonymization of pre-training and fine-tuning data. These results demonstrate that pseudonymizing training data to reduce privacy risks can be done without harming data utility for training PLMs.

Fault vibration model driven fault-aware domain generalization framework for bearing fault diagnosis

Deep learning methods can learn effective representations from the data, simplifying the fault diagnosis process and improving accuracy. However, the lack of data presents a significant challenge to effectively implementing fault diagnosis driven by deep learning. Simulation models can generate rich source domain data, which can be combined with domain adaptation methods to realize fault diagnosis tasks under sample scarcity. Nevertheless, there are some problems with existing simulation-driven fault diagnosis frameworks: (1) The choice of parameters for simulation models is usually empirical, which can lead to domain-specific components differences between the simulation data and the actual data. (2) It is difficult for the simulation models to generate the background noise of the real equipment, which will cause fault information to be covered, severely limiting the robustness of the fault diagnosis network. (3) Studies based on domain adaptation methods require a certain amount of target domain data for training, which is difficult to meet in real industrial production. To address these issues, this paper proposes a fault-aware domain generalization bearing fault diagnosis framework driven by fault vibration model, an analytical simulation model. Firstly, a spectral background estimation algorithm based on adaptive clutter separation (ACS) and maximum spectral envelope (MSE), termed ACS-MSE, is designed to estimate the potential resonance frequencies in the real system, guiding the parameters setting of the fault diagnosis vibration model and ensuring consistency between the simulated data and the actual data in the label space. Secondly, the normal actual signal is combined with the fault simulation signal to create a mixed signal that considers real background noise, and a fault-aware autoencoder is developed to extract the fault-related impact components covered by the background noise. Finally, a multi resonance frequencies domain generalization fault diagnosis network based on contrastive learning is designed to extract generalized fault features crossing different resonance frequencies and realize diagnosis tasks without real fault data. Experimental results on an open dataset and a custom dataset showed that the proposed method achieves higher accuracy compared to some advanced domain generalization and even domain adaptation methods. Ablation studies verify the importance and irreplaceability of the individual components of the proposed method.

WSPolyp-SAM: Weakly Supervised and Self-Guided Fine-Tuning of SAM for Colonoscopy Polyp Segmentation

Ensuring precise segmentation of colorectal polyps holds critical importance in the early diagnosis and treatment of colorectal cancer. Nevertheless, existing deep learning-based segmentation methods are fully supervised, requiring extensive, precise, manual pixel-level annotation data, which leads to high annotation costs. Additionally, it remains challenging to train large-scale segmentation models when confronted with limited colonoscopy data. To address these issues, we introduce the general segmentation foundation model—the Segment Anything Model (SAM)—into the field of medical image segmentation. Fine-tuning the foundation model is an effective approach to tackle sample scarcity. However, current SAM fine-tuning techniques still rely on precise annotations. To overcome this limitation, we propose WSPolyp-SAM, a novel weakly supervised approach for colonoscopy polyp segmentation. WSPolyp-SAM utilizes weak annotations to guide SAM in generating segmentation masks, which are then treated as pseudo-labels to guide the fine-tuning of SAM, thereby reducing the dependence on precise annotation data. To improve the reliability and accuracy of pseudo-labels, we have designed a series of enhancement strategies to improve the quality of pseudo-labels and mitigate the negative impact of low-quality pseudo-labels. Experimental results on five medical image datasets demonstrate that WSPolyp-SAM outperforms current fully supervised mainstream polyp segmentation networks on the Kvasir-SEG, ColonDB, CVC-300, and ETIS datasets. Furthermore, by using different amounts of training data in weakly supervised and fully supervised experiments, it is found that weakly supervised fine-tuning can save 70% to 73% of annotation time costs compared to fully supervised fine-tuning. This study provides a new perspective on the combination of weakly supervised learning and SAM models, significantly reducing annotation time and offering insights for further development in the field of colonoscopy polyp segmentation.

Poster Design Research Based on Deep Learning Automatic Image Generation Algorithm

Although this research has made significant progress in image generation models, they still face issues such as insufficient diversity of generated images, poor quality of high-resolution images, and the need for a large amount of training data for model optimization. This paper studies poster design based on deep learning automatic image generation algorithm, using a recursive supervised image generation algorithm framework of generative adversarial networks for multi-view image generation and super-resolution generation tasks of small sample digital poster images. Various improvements have been proposed to enhance the performance of the GAN network model for poster design image generation tasks. Based on experimental research, this paper’s model uses generative adversarial networks to distinguish randomly cropped low resolution and high-resolution poster blocks, ensuring that high-resolution posters maintain their original resolution canvas texture and brush strokes, effectively improving the automatic generation effect of poster images. The evaluation results show that the quantitative evaluation of the proposed algorithm model in knowledge management is distributed in a reasonable range, which indicates that the proposed algorithm model has good performance in knowledge management. The poster design model based on deep learning automatic image generation algorithm proposed in this paper has certain effects. In subsequent practice, the automatic image generation algorithm can be combined with practical needs to improve the efficiency and design effect of poster design.

A Novel Approach to Surface Roughness Virtual Sample Generation to Address the Small Sample Size Problem in Ultra-Precision Machining.

Surface roughness is one of the main bases for measuring the surface quality of machined parts. A large amount of training data can effectively improve model prediction accuracy. However, obtaining a large and complete surface roughness sample dataset during the ultra-precision machining process is a challenging task. In this article, a novel virtual sample generation scheme (PSOVSGBLS) for surface roughness is designed to address the small sample problem in ultra-precision machining, which utilizes a particle swarm optimization algorithm combined with a broad learning system to generate virtual samples, enriching the diversity of samples by filling the information gaps between the original small samples. Finally, a set of ultra-precision micro-groove cutting experiments was carried out to verify the feasibility of the proposed virtual sample generation scheme, and the results show that the prediction error of the surface roughness prediction model was significantly reduced after adding virtual samples.

Proto-Adapter: Efficient Training-Free CLIP-Adapter for Few-Shot Image Classification.

Large vision-language models, such as Contrastive Vision-Language Pre-training (CLIP), pre-trained on large-scale image-text datasets, have demonstrated robust zero-shot transfer capabilities across various downstream tasks. To further enhance the few-shot recognition performance of CLIP, Tip-Adapter augments the CLIP model with an adapter that incorporates a key-value cache model constructed from the few-shot training set. This approach enables training-free adaptation and has shown significant improvements in few-shot recognition, especially with additional fine-tuning. However, the size of the adapter increases in proportion to the number of training samples, making it difficult to deploy in practical applications. In this paper, we propose a novel CLIP adaptation method, named Proto-Adapter, which employs a single-layer adapter of constant size regardless of the amount of training data and even outperforms Tip-Adapter. Proto-Adapter constructs the adapter's weights based on prototype representations for each class. By aggregating the features of the training samples, it successfully reduces the size of the adapter without compromising performance. Moreover, the performance of the model can be further enhanced by fine-tuning the adapter's weights using a distance margin penalty, which imposes additional inter-class discrepancy to the output logits. We posit that this training scheme allows us to obtain a model with a discriminative decision boundary even when trained with a limited amount of data. We demonstrate the effectiveness of the proposed method through extensive experiments of few-shot classification on diverse datasets.

Explainable proactive control of industrial processes

One of the goals of Industry 4.0 is the adoption of data-driven models to enhance various aspects of the manufacturing process, such as monitoring equipment conditions, ensuring product quality, detecting failures, and preparing optimal maintenance plans. However, many machine-learning algorithms require a large amount of training data to reach desired performance. In numerous industrial applications, such data is either not available or its acquisition is a costly process. In such cases, simulation frameworks are employed to replicate the behavior of real-world facilities and generate data for further analysis. Simulation frameworks typically provide high-quality data but are often slow which can be problematic when real-time decision-making is required. Control approaches based on simulation-based data commonly face challenges related to inflexibility, particularly in dynamic production environments undergoing frequent reconfiguration and upgrades. This paper introduces a method that seeks to strike a balance between the reliance on simulated data and the limited robustness of simulation-based control methods. This balance is achieved by supplementing available data with additional expert knowledge, enabling the matching of similar data sources and their combination for reuse. Furthermore, we augment the methods with an explainability layer, facilitating collaboration between the human expert and the AI system, leading to informed and actionable decisions. The performance of the proposed solution is demonstrated through a case study on gas production from an underground reservoir resulting in reduced downtime, heightened process reliability, and enhanced overall performance. This paper builds upon our conference paper (Kuk et al., 2023), addressing the same problem with an extended, more generic methodology, and presenting entirely new results.

Meta-learning Achieves High Accuracy with a Small Amount of Training Data

Meta-learning Achieves High Accuracy with a Small Amount of Training Data

Towards an optimization of automatic defect detection by artificial neural network using Lamb waves

This paper presents a damage detection method based on the inverse pattern recognition technique by artificial neural network (ANN) using ultrasonic waves. Lamb waves are guided elastic waves, are widely employed in nondestructive testing thanks to their attractive properties such as their sensitivity to the small defects. In this work, finite element method was conducted by Abaqus to study Lamb modes propagation. A data collection is performed by the signals recorded from the sensor of 300 models: healthy and damaged plates excited by a tone burst signal with the frequencies: 100 kHz, 125 kHz, 150 kHz, 175 kHz, 200 kHz, and 225 kHz. The captured signals in undamaged plat are the baseline, whereas the signals measured in damaged plates are recorded for various positions of external rectangular defects. To reduce the amount of training data, only two peaks of measured signals are required to be the input of the model. Continuous wavelet transform (CWT) was adopted to calculate the key features of the signal in the time domain. The feed forward neural network is implemented using MATLAB program. The data are divided as follows: 70% for training the model, 25% for the validation, and 5% for the test. The proposed model is accurate estimating the position of the defect with an accuracy of 99.98%.

Federated Learning and Unlearning as Enablers of Wind Turbine Digital Twins

Real-time capable models are paramount for the successful adaptation of digital twin technology into industries such as wind energy, but high-fidelity physics-based models cannot achieve the required speed, while data-driven and hybrid methods require large amounts of training data which is typically confidential. In this work, the combination of federated learning with hybrid modeling is proposed to train fast and reliable models across multiple confidential data sets owned by different stakeholders. The approach is demonstrated on physics-guided neural networks to estimate the lift and drag of wind turbine airfoils. A scenario is devised where multiple confidential data subsets are confined to different client devices. It is shown that the physics-guided neural networks can be trained through federated learning across those devices and data subsets and that the resulting models can recover and even surpass the accuracy of a model that is trained conventionally by merging the data subsets on a single device. The presented approach is highly scalable and can be easily adapted to many other applications. This work also discusses federated unlearning methods, which allow data owners to remove all traces of a data subset used in training if they decide to revoke their contribution.

GAN Based Multi-Class Skin Disease Classification: Deep Learning Approach

Abstract: Skin diseases pose significant diagnostic and treatment challenges due to their diverse and complex manifestations. Convolutional neural networks (CNNs) have demonstrated superior capabilities in image classification tasks, including skin disease identification. However, the performance of CNN models depends heavily on the quality and quantity of training data, which often suffers from limitations such as imbalance and sparsity. This project proposes an approach new approach to address these challenges by integrating generative adversarial networks (GANs). ) with CNN for multi-class skin disease classification. The GAN-based system aims to improve the diversity and quantity of the training dataset by generating synthetic images of various skin conditions. Through an adversarial training process, the generator network learns to generate realistic images of skin diseases, while the discriminator network distinguishes between real and synthetic data. Figures The synthetic images generated by the GAN are then combined with the real dataset to train the model CNN. specially designed to classify skin diseases.

Two-dimensional hybrid incremental learning (2DHIL) framework for semantic segmentation of skin tissues

This study aims to enhance the robustness and generalization capability of a deep learning transformer model used for segmenting skin carcinomas and tissues through the introduction of incremental learning. Deep learning AI models demonstrate their claimed performance only for tasks and data types for which they are specifically trained. Their performance is severely challenged for the test cases which are not similar to training data thus questioning their robustness and ability to generalize. Moreover, these models require an enormous amount of annotated data for training to achieve desired performance. The availability of large annotated data, particularly for medical applications, is itself a challenge. Despite efforts to alleviate this limitation through techniques like data augmentation, transfer learning, and few-shot training, the challenge persists. To address this, we propose refining the models incrementally as new classes are discovered and more data becomes available, emulating the human learning process. However, deep learning models face the challenge of catastrophic forgetting during incremental training. Therefore, we introduce a two-dimensional hybrid incremental learning framework for segmenting non-melanoma skin cancers and tissues from histopathology images. Our approach involves progressively adding new classes and introducing data of varying specifications to introduce adaptability in the models. We also employ a combination of loss functions to facilitate new learning and mitigate catastrophic forgetting. Our extended experiments demonstrate significant improvements, with an F1 score reaching 91.78, mIoU of 93.00, and an average accuracy of 95%. These findings highlight the effectiveness of our incremental learning strategy in enhancing the robustness and generalization of deep learning segmentation models while mitigating catastrophic forgetting.

A geometric algebra-based approach for myoelectric pattern recognition control and faster prosthesis recalibration

ObjectiveAlthough many advancements have been made on myoelectric pattern-recognition, the control of poly-articulated upper-limb prostheses remains insufficiently robust. Electrode-shift, sweat or fatigue degrade the performance of classifiers over time, resulting in unfruitful device usage and frequent re-calibration. To tackle this issue, here we introduce two models − µP6 and µP8 − that combine Geometric Algebra with nearest-neighbor classification. We aim at reducing both the necessary training data and training time and, unlike most current state-of-the-art algorithms, we exploit an alternative geometric representation (and visualization) of the EMG signal as different polygons for different types of gestures, facilitating the explanation of the decision-making process to a layman. Moreover, we explore four abstention strategies to reduce the number of misclassifications. MethodsWe perform an offline analysis on two datasets, alongside two other standard models: nonlinear logistic regression (NLR) and linear discriminant analysis (LDA). ResultsEven with few training data, the proposed algorithms achieve high F1-scores (>0.95), significantly higher or non-significantly different from the values obtained with NLR and LDA, while maintaining relatively low abstentions rates and training times (<2 ms). ConclusionThe proposed algorithms allow to reduce the amount of training data and training times without compromising recognition rates. SignificanceThe proposed algorithms may contribute for a faster prosthesis re-calibration procedure while allowing to re-gain high recognition rates. Furthermore, the decision-making process is explainable and interpretable, potentially improving user trust and acceptance.

Development of Specialization Selection on The Job Training for Human Resources in The Hospitality Industry through E-OJTKU

The previous problem in this research was exploring the potential of students in choosing departments in hotels, namely front office, housekeeping, and food and beverage to carry out internships. The need for this selection recommendation is based not only on student desires or the availability of opportunities to learn to work in the existing hotel industry but also based on recommended grades from semesters 1 to 3 which fulfill the requirements for industrial internships. This research uses a qualitative descriptive method with a case study approach. Data collection techniques through observation, interviews and document study. Interviews use a semi-structured model so that answers are open-ended and can be developed for in-depth analysis. E-ojtku is an innovation resulting from research based website and android to improve the competency of Hospitality Management students who received recognition for 40 credits for OJT 1 and OJT 2 (1 year in industry / Hotels) for a career in Hotels. The test results using e-ojtku obtained good results with an accuracy of up to 94.44% with a ratio of 80% training data and 20% test data, the ratio of the amount of training data can affect the accuracy of the values in each experiment.