General Neural Network Research Articles

Remote sensing-enhanced transfer learning approach for agricultural damage and change detection: A deep learning perspective

With the continuous advancement of science and technology, there has been a growing awareness of safety among people worldwide. Natural disasters such as wildfires, earthquakes, and floods pose persistent threats to both lives and property on our planet, which serves as our fundamental habitat. While it is impossible to prevent or entirely avert these calamities, rapid identification of affected areas and prompt damage assessment post-disaster can significantly aid in the formulation of effective rescue strategies, ultimately saving more lives. This article delves into the application of transfer learning in satellite image damage assessment—a methodology that involves transferring previously acquired knowledge to enhance a model's adaptability to new tasks. Given the limited availability of datasets for satellite image analysis, transfer learning proves to be an effective approach. Specifically, the study proposes a transfer learning method based on YOLOv5 for satellite image damage assessment. Initially, a general convolutional neural network model is trained using a substantial dataset of natural images. Subsequently, the early layers of this model are frozen, while the later layers undergo training to adapt to satellite image data. Fine-tuning is then employed to further enhance the overall model performance. The results demonstrate that this approach yields a high accuracy rate in satellite image damage assessment. Moreover, compared to conventional deep learning methods, the proposed method effectively leverages pre-trained models' knowledge, thereby reducing data dependency. Additionally, it displays robust generalization capabilities across diverse tasks and datasets, underscoring its potential for facilitating transfer learning across various domains.

Fuzzy adaptive event-triggered synchronization control mechanism for T–S fuzzy RDNNs under deception attacks

In this paper, a fuzzy-dependent adaptive event-triggered mechanism (FAETM) for synchronizing Takagi–Sugeno (T–S) fuzzy reaction–diffusion neural networks (RDNNs) is developed while considering deception attacks. Firstly, a general neural network model considering both fuzzy logic rules and reaction–diffusion terms is established. Secondly, a FAETM based on an aperiodic sampling period is presented under deception attacks to alleviate the communication burden, wherein the adaptive threshold function is dependent on membership functions. Moreover, a membership-function-dependent asymmetric Lyapunov functional (LF) is constructed, and some positive-definite and symmetric constraints of matrices are removed as a result. Based on the LF method and integral inequality techniques, the H∞ synchronization criteria for the T–S fuzzy RDNNs are presented by linear matrix inequalities (LMIs). Finally, one simulation example is exploited to demonstrate the feasibility and validity of the established method, and the outcome is applied to image secure communication.

Design of a novel hybrid quantum deep neural network in INEQR images classification

We redesign the parameterized quantum circuit in the quantum deep neural network, construct a three-layer structure as the hidden layer, and then use classical optimization algorithms to train the parameterized quantum circuit, thereby propose a novel hybrid quantum deep neural network (HQDNN) used for image classification. After bilinear interpolation reduces the original image to a suitable size, an improved novel enhanced quantum representation (INEQR) is used to encode it into quantum states as the input of the HQDNN. Multi-layer parameterized quantum circuits are used as the main structure to implement feature extraction and classification. The output results of parameterized quantum circuits are converted into classical data through quantum measurements and then optimized on a classical computer. To verify the performance of the HQDNN, we conduct binary classification and three classification experiments on the MNIST (Modified National Institute of Standards and Technology) data set. In the first binary classification, the accuracy of 0 and 4 exceeds 98%. Then we compare the performance of three classification with other algorithms, the results on two datasets show that the classification accuracy is higher than that of quantum deep neural network and general quantum convolutional neural network.

Tongue feature dataset construction and real-time detection.

Tongue diagnosis in traditional Chinese medicine (TCM) provides clinically important, objective evidence from direct observation of specific features that assist with diagnosis. However, the current interpretation of tongue features requires a significant amount of manpower and time. TCM physicians may have different interpretations of features displayed by the same tongue. An automated interpretation system that interprets tongue features would expedite the interpretation process and yield more consistent results. This study applied deep learning visualization to tongue diagnosis. After collecting tongue images and corresponding interpretation reports by TCM physicians in a single teaching hospital, various tongue features such as fissures, tooth marks, and different types of coatings were annotated manually with rectangles. These annotated data and images were used to train a deep learning object detection model. Upon completion of training, the position of each tongue feature was dynamically marked. A large high-quality manually annotated tongue feature dataset was constructed and analyzed. A detection model was trained with average precision (AP) 47.67%, 58.94%, 71.25% and 59.78% for fissures, tooth marks, thick and yellow coatings, respectively. At over 40 frames per second on a NVIDIA GeForce GTX 1060, the model was capable of detecting tongue features from any viewpoint in real time. This study constructed a tongue feature dataset and trained a deep learning object detection model to locate tongue features in real time. The model provided interpretability and intuitiveness that are often lacking in general neural network models and implies good feasibility for clinical application.

Operational transfer path analysis based on neural network

Noise, vibration, and harshness (NVH) issues induced by vibrating source to receivers are sensitive concerns in the automotive industry, significantly determining the product quality. Transfer path analysis (TPA) serves as a crucial tool in addressing these concerns. However, classical and component-based TPA methods necessitate the measurement of all transfer functions associated with the NVH phenomenon, causing high experimental costs; moreover, operational TPA (OTPA) involves matrix inversion processes that may result in the loss of important information. In this study, a neural network-based OTPA (NOTPA) method is proposed that eliminates the need for a matrix inversion process and utilizes only operational measurement. To derive the physical meaning of transmissibility obtained from the trained neural network model, the learning algorithm for NOTPA was developed, and appropriate learning rules were applied to the model. In the general neural network model, training was performed using real-valued parameters. However, the phase of the signal is crucial information in the theoretical formulation of TPA; thus, a complex-valued propagation algorithm was established. Furthermore, an appropriate activation function was chosen to derive the actual transfer mechanism from the trained complex-valued parameters. To experimentally verify the NOTPA method, a testbed resembling automotive vehicle structure was constructed. Estimation of sound pressure and noise contribution analysis were conducted using the NOTPA method and compared with conventional TPA methods. The proposed method demonstrated higher accuracy compared to the conventional OTPA method and successfully identified the main transfer path compared to the classical TPA. Furthermore, by comparing the estimated sound pressure according to the architecture of the model, the consistency of the proposed model was verified.

Simulation and prediction of water temperature in a water transfer channel during winter periods using a new approach based on the wavelet noise reduction-deep learning method

Abstract In winter, the water transfer channel of the Middle Route of South-to-North Water Transfer Project (MR-StNWTP) in China always encounters ice problems. The preciously simulation and prediction of water temperature is essential for analyzing the ice condition, which is important for the safety control of the water transfer channel in winter. Due to the difference of specific heat between water and air, when the air temperature rises and falls dramatically, the range of change of water temperature is relatively small and has a lag, which often affects the accuracy of simulation and prediction of water temperature based on air temperature. In the present study, a new approach for simulating and predicting water temperature in water transfer channels in winter has been proposed. By coupling the neural network theory to equations describing water temperature, a model has been developed for predicting water temperature. The temperature data of prototype observations in winter are preprocessed through the wavelet decomposition and noise reduction. Then, the wavelet soft threshold denoising method is used to eliminate the fluctuation of certain temperature data of prototype observations, and the corresponding water temperature is calculated afterward. Compared to calculation results using both general neural network and multiple regression approaches, the calculation results using the proposed model agree well with those of prototype measurements and can effectively improve the accuracy of prediction of water temperature.

Modeling of the biosorption of Fe(III) ions by olive‐stone activated carbon

This study deals with the optimization of Fe(III) ion removal using activated carbon from olive stone waste using advanced machine learning models. The main objective is to evaluate and compare the performance of machine learning models, specifically multilayer perceptron artificial neural network (MLP‐ANN), general regression artificial neural network (GR‐ANN), radial basis function artificial neural network (RBF‐ANN), and particle swarm optimization artificial neural network (PSO‐ANN) in predicting Fe(III) removal efficiency. Experimental data on adsorption parameters were used to train and test the models. Techniques such as tuning hidden layer neurons, optimizing propagation values, and using a Taguchi approach PSO algorithm were applied to improve the models. For the MLP‐ANN model, the optimal configuration contains 13 neurons in the hidden layer. Concerning the parameters involved in the PSO‐ANN model, the coefficient C2 and the particle have the main effect on the reduction of the error. Their contributions are respectively 49% and 19%. The PSO‐ANN model showed superior performance with the highest regression coefficient (0.9997) and remarkable prediction accuracy, surpassing other models such as MLP‐ANN and GR‐ANN. This research suggests that innovative optimization techniques, particularly using PSO algorithms, significantly enhance the predictive capabilities of machine learning models in complex adsorption processes, contributing to more accurate Fe(III) removal models.

Grounded language acquisition through the eyes and ears of a single child.

Starting around 6 to 9 months of age, children begin acquiring their first words, linking spoken words to their visual counterparts. How much of this knowledge is learnable from sensory input with relatively generic learning mechanisms, and how much requires stronger inductive biases? Using longitudinal head-mounted camera recordings from one child aged 6 to 25 months, we trained a relatively generic neural network on 61 hours of correlated visual-linguistic data streams, learning feature-based representations and cross-modal associations. Our model acquires many word-referent mappings present in the child's everyday experience, enables zero-shot generalization to new visual referents, and aligns its visual and linguistic conceptual systems. These results show how critical aspects of grounded word meaning are learnable through joint representation and associative learning from one child's input.

Distilling Invariant Representations With Domain Adversarial Learning for Cross-Subject Children Seizure Prediction

Seizure prediction based on electroencephalogram (EEG) has great potential to improve patients’ life quality. Due to the high heterogeneity in distributions of EEG signals among different patients, conventional studies usually show poor generalization ability when transferring the model to new patients, which also leads to difficulties in clinical applications. To alleviate the challenging issue concerning cross-subject domain shift, we propose a transformer-based domain adversarial model. Our model first adopts a pre-trained general neural network to extract common features from the EEG signals of available patients. Then, we design a distiller module and a domain discriminator module to perform domain adaptation training based on a small amount of labeled data from the new-coming patient. During adaptation process, conditional domain adversarial training with the addition of label information is employed to remove patient-related information from the extracted features to learn a common seizure feature space among different patients. Our proposed seizure prediction method is evaluated on the CHB-MIT EEG database. The proposed model achieves a sensitivity of 79.5%, a false alarm rate (FPR) of 0.258/h, and an AUC of 0.814. Experimental results demonstrate that the proposed method can effectively reduce inter-patient domain disparity compared to state-of-the-art methods.

Prediction of Converter Tapping Weight Based on Principal Component Analysis–Whale Optimization Algorithm–Backpropagation Algorithm Neural Network Model

Tapping weight is an important parameter in converter blowing process, wherein precisely predicting the quantity of steel required in an alloy baking converter can effectively guide the requirement of alloy ingredients. In practical production, the main approach is empirical estimation, despite its low accuracy. Employing a general neural network model for prediction requires to gather the converter blowing parameters and endpoint temperature measurement sampling parameters as the model inputs. However, this data cannot be obtained until the blowing process reaches its endpoint, rendering it impractical for alloy batching that requires advance preparation for baking. In this study, a principal component analysis–whale optimization algorithm–backpropagation algorithm (PCA–WOA–BP) neural network tapping prediction model is developed using raw material parameters available before alloy baking as input. This model is integrated into the intelligent alloy reduction model of a factory. The model achieves a regression coefficient R2 of 0.823, with 98.53% of furnaces having a prediction error of less than 2 t. The tapping weight of 20 consecutive heats in actual production is predicted; the error range is less than 80% within 1 t, and the converter tapping weight is predicted accurately.

Biogas production prediction model of food waste anaerobic digestion for energy optimization using mixup data augmentation-based global attention mechanism.

Achieving rapid, efficient, and cost-effective anaerobic digestion (AD) of food waste is a key means to improve the efficiency of food waste treatment. However, in view of the shortage of historical anaerobic digestion data, the limitation of general neural networks in predicting biogas production, and its sensitivity to abnormal variation points, achieving accurate prediction of biogas production is not easy. This paper proposes a novel biogas production prediction model of food waste AD for energy optimization based on the mixup data augmentation integrating an improved global attention mechanism long short-term memory (LSTM). Taking the AD data of the actual factory as samples, the mixup data augmentation is introduced to generate virtual samples with the similar distribution as original samples. Then original samples and generated virtual samples are used as the input of the global attention mechanism LSTM to establish the food waste AD biogas production prediction model. Finally, the proposed method is applied in the biogas production prediction of actual food waste treatment plants. Compared with other industrial modeling models, the experimental results show that the proposed method has the highest prediction accuracy of 0.988, which performs well in predicting biogas production and can effectively guide and timely adjust feed configuration of AD plants.

Hyperbolic Binary Neural Network.

Binary neural network (BNN) converts full-precision weights and activations into their extreme 1-bit counterparts, making it particularly suitable for deployment on lightweight mobile devices. While BNNs are typically formulated as a constrained optimization problem and optimized in the binarized space, general neural networks are formulated as an unconstrained optimization problem and optimized in the continuous space. This article introduces the hyperbolic BNN (HBNN) by leveraging the framework of hyperbolic geometry to optimize the constrained problem. Specifically, we transform the constrained problem in hyperbolic space into an unconstrained one in Euclidean space using the Riemannian exponential map. On the other hand, we also propose the exponential parametrization cluster (EPC) method, which, compared with the Riemannian exponential map, shrinks the segment domain based on a diffeomorphism. This approach increases the probability of weight flips, thereby maximizing the information gain in BNNs. Experimental results on CIFAR10, CIFAR100, and ImageNet classification datasets with VGGsmall, ResNet18, and ResNet34 models illustrate the superior performance of our HBNN over state-of-the-art methods.

Sign Language Recognition Using Graph and General Deep Neural Network Based on Large Scale Dataset

Sign Language Recognition Using Graph and General Deep Neural Network Based on Large Scale Dataset

Data augmentation for 3D seismic fault interpretation using deep learning

Manual seismic interpretation of faults is a tedious and complicated process, which is prone to human error and bias. A semi-automatic approach for interpreting faults on seismic data is to use attributes that highlight discontinuities. These methods need to be optimized by the interpreter, thus constantly rely on the interpreter's knowledge. Recently, Machine Learning (ML) techniques in general and Convolutional Neural Networks (CNN) as part of Deep Neural Networks (DNN), in particular, have been used to detect and image faults on seismic data in order to make the process more automated and accurate. Supervised CNN learns and evolves from manually annotated or labeled fault interpretations.In this study, we have applied supervised 2D CNN to image faults on seismic data through binary segmentation. The study was performed on 3D marine seismic surveys in off-shore Norway collected from three separate locations along the Norwegian Continental Shelf, utilizing Efficient UNET CNN architecture. We applied data augmentation (geometric transformations) and hyperparameter tuning to improve the learning process and performance of the deep learning algorithm.We used noise content, acquisition type, imaging type, and fault scale as the main criteria to select seismic surveys. The application of data augmentation to the training and testing data generally led to improvement in the performance of 2D CNN on fault predictions, although the amount of improvement varied with respect to different surveys. The initial 2D CNN fault prediction improvement mainly relied on the quality, and size of faults present in the 3D seismic volumes. Further, improvement was achieved by the adjustment of certain hyperparameters affecting the training and testing process of the 2D CNN. However, we found little to no improvement on one seismic volume containing high levels of noise.

Improved Transformer for Time Series Senescence Root Recognition.

The root is an important organ for plants to obtain nutrients and water, and its phenotypic characteristics are closely related to its functions. Deep-learning-based high-throughput insitu root senescence feature extraction has not yet been published. In light of this, this paper suggests a technique based on the transformer neural network for retrieving cotton's insitu root senescence properties. High-resolution insitu root pictures with various levels of senescence are the main subject of the investigation. By comparing the semantic segmentation of the root system by general convolutional neural networks and transformer neural networks, SegFormer-UN (large) achieves the optimal evaluation metrics with mIoU, mRecall, mPrecision, and mF1 metric values of 81.52%, 86.87%, 90.98%, and 88.81%, respectively. The segmentation results indicate more accurate predictions at the connections of root systems in the segmented images. In contrast to 2 algorithms for cotton root senescence extraction based on deep learning and image processing, the insitu root senescence recognition algorithm using the SegFormer-UN model has a parameter count of 5.81 million and operates at a fast speed, approximately 4 min per image. It can accurately identify senescence roots in the image. We propose that the SegFormer-UN model can rapidly and nondestructively identify senescence root in insitu root images, providing important methodological support for efficient crop senescence research.

A general graph neural network based implicit solvation model for organic molecules in water.

The dynamical behavior of small molecules in their environment can be studied with classical molecular dynamics (MD) simulations to gain deeper insight on an atomic level and thus complement and rationalize the interpretation of experimental findings. Such approaches are of great value in various areas of research, e.g., in the development of new therapeutics. The accurate description of solvation effects in such simulations is thereby key and has in consequence been an active field of research since the introduction of MD. So far, the most accurate approaches involve computationally expensive explicit solvent simulations, while widely applied models using an implicit solvent description suffer from reduced accuracy. Recently, machine learning (ML) approaches that provide a probabilistic representation of solvation effects have been proposed as potential alternatives. However, the associated computational costs and minimal or lack of transferability render them unusable in practice. Here, we report the first example of a transferable ML-based implicit solvent model trained on a diverse set of 3 000 000 molecular structures that can be applied to organic small molecules for simulations in water. Extensive testing against reference calculations demonstrated that the model delivers on par accuracy with explicit solvent simulations while providing an up to 18-fold increase in sampling rate.

A diabatization method based upon integrating the diabatic potential gradient difference.

In this work we develop a new scheme to construct a diabatic potential energy matrix (DPEM). We propose a diabatization method which is based on integrating the diabatic potential gradient difference to diabatize adiabatic ab initio energies. This method is capable of performing high-precision adiabatic-to-diabatic transformations, with a unique advantage in effectively handling the significant fluctuations in derivative-couplings caused by conical intersection (CI) seams. The above scheme is applied to the DPEM construction of the Na(3p) + H2 → NaH + H reaction. The fitting data including adiabatic energies, energy gradients and derivative-couplings obtained from a previous benchmark DPEM are diabatized and fitted using a general neural network fitting procedure to generate the DPEM. The produced DPEM can effectively describe nonadiabatic processes involving different electronic states. We further perform quantum dynamical calculations on the new DPEM and the previous benchmark DPEM, and the obtained results demonstrate the effectiveness of our scheme.

Breast morphology classification method based on generalized neural network

In order to establish a breast shape classification method with general indicators, this paper selects six measurement indicators that can characterize the three-dimensional shape of the breast, such as circumference, length, width, and height. The relevant data of the measurement indicators are used to establish a generalized neural network model, and the generalized neural network is applied to the evaluation of breast shape classification recognition. The MATLAB software toolbox is used, based on training sample data, We have established a generalized neural network (GRNN) model, which involves establishing corresponding functional relationships through training to approximate the training data. After obtaining the function relationship through training, the predicted values were obtained, and the prediction results reflected that the established generalized neural network GRNN model is quite reliable for predicting the category of breast shape. By establishing a mapping relationship between the input quantity of breast feature parameters and the output quantity of categories, training samples can be established. When encountering new test data, these data can also be used as test samples. Based on the established model, appropriate spreads can be adjusted to predict the output quantity, providing relevant personnel with basic and simple basis for judgment and evaluation. It can also provide convenience for customers in various consumption channels Accurate bra size recommendations can help consumers purchase bras that fit well and are comfortable.

A neural network finite element method for contact mechanics

Computational biomechanical models of organ systems utilizing the finite element (FE) method have been extensively applied to the study of normal and pathophysiological function. While a robust method and providing for many unique pathophysiological insights, traditional FE methods remain prohibitively slow for real-time many-query clinical applications. To meet these demanding computational requirements, we have developed a general neural network finite element (NNFE) approach for hyperelastic soft tissue organ simulations that can produce equivalent accuracy within clinically relevant time frames. However, many organ systems involve contact between structural elements (e.g. heart valve leaflets), which has not been previously addressed in the NNFE and related approaches. In the present work, we developed a NNFE-based method for contact between elastic deformable bodies. We exploited the fact that stable equilibrium solutions in nonlinear hyperelasticity minimize the potential energy to train the neural network, with contact incorporated through an potential energy penalty. This penalty was discretized and computationally implemented in JAX (Google, Inc.) in a way that could be statically allocated and jit compiled for rapid computation on modern GPU hardware. Following our previous NNFE formulation, we represented the problem domain geometry and enforced the necessary boundary conditions using a Non-Uniform Rational B-Splines (NURBS) based geometry scheme. NURBS were chosen for their ability to smoothly represent body and surface geometries. To demonstrate the method, we developed and verified two basic contact problems using a hyperelastic material model: indentation and flap. Both problems were simulated with a fully connected neural network with a parallel linear layer. Training continued until the gradient of the energy with respect to network parameters dropped below 10−2, taking less than six hours in each case studied. Once trained, the NNFE approach was capable of accurately predicting the equilibrium configurations of representative contact problems. Moreover, the NNFE contact approach required nearly six order of magnitude less time to evaluate than an equivalent FEniCSx implementation that utilized a similar mesh of cubic order Lagrange elements (0.01 s compared to 1715 seconds, respectively). Moreover, the solutions were trained over full ranges of loading and were thus able to faithfully represent families of solutions without requiring retraining. While addressing basic problems in contact, this study serves as crucial stepping stone to generate NNFE organ simulations for rapid prediction of patient specific solutions that involve contact.

Radiation-induced acoustic signal denoising using a supervised deep learning framework for imaging and therapy monitoring.

Radiation-induced acoustic (RA) imaging is a promising technique for visualizing the invisible radiation energy deposition in tissues, enabling new imaging modalities and real-time therapy monitoring. However, RA imaging signal often suffers from poor signal-to-noise ratios (SNRs), thus requiring measuring hundreds or even thousands of frames for averaging to achieve satisfactory quality. This repetitive measurement increases ionizing radiation dose and degrades the temporal resolution of RA imaging, limiting its clinical utility. In this study, we developed a general deep inception convolutional neural network (GDI-CNN) to denoise RA signals to substantially reduce the number of frames needed for averaging. The network employs convolutions with multiple dilations in each inception block, allowing it to encode and decode signal features with varying temporal characteristics. This design generalizes GDI-CNN to denoise acoustic signals resulting from different radiation sources. The performance of the proposed method was evaluated using experimental data of x-ray-induced acoustic, protoacoustic, and electroacoustic signals both qualitatively and quantitatively. Results demonstrated the effectiveness of GDI-CNN: it achieved x-ray-induced acoustic image quality comparable to 750-frame-averaged results using only 10-frame-averaged measurements, reducing the imaging dose of x-ray-acoustic computed tomography (XACT) by 98.7%; it realized proton range accuracy parallel to 1500-frame-averaged results using only 20-frame-averaged measurements, improving the range verification frequency in proton therapy from 0.5 to 37.5 Hz; it reached electroacoustic image quality comparable to 750-frame-averaged results using only a single frame signal, increasing the electric field monitoring frequency from 1 fps to 1k fps. Compared to lowpass filter-based denoising, the proposed method demonstrated considerably lower mean-squared-errors, higher peak-SNR, and higher structural similarities with respect to the corresponding high-frame-averaged measurements. The proposed deep learning-based denoising framework is a generalized method for few-frame-averaged acoustic signal denoising, which significantly improves the RA imaging's clinical utilities for low-dose imaging and real-time therapy monitoring.