Predictive health behavior modeling using multimodal feature correlations via Medical Internet-of-Things devices.

Due to the advent of IoT (Internet of Things) based devices that help to monitor different human behavioral aspects. These aspects include sleeping patterns, activity patterns, heart rate variability (HRV) patterns, location-based moving patterns, blood oxygen levels, etc. A correlative study of these patterns can be used to find linkages of behavioral patterns with human health conditions. To perform this task, a wide variety of models is proposed by researchers, but most of them vary in terms of used parameters, which limits their accuracy of analysis. Moreover, most of these models are highly complex and have lower parameter flexibility, thus, cannot be scaled for real-time use cases. To overcome these issues, this paper proposes design of a behavior modeling method that assists in future health predictions via multimodal feature correlations using medical IoT devices via deep transfer learning analysis. The proposed model initially collects large-scale sensor data about the subjects, and correlates them with the existing medical conditions. This correlation is done via extraction of multidomain feature sets that assist in spectral analysis, entropy evaluations, scaling estimation, and window-based analysis. These multidomain feature sets are selected by a Firefly Optimizer (FFO) and are used to train a Recurrent Neural Network (RNN) Model, that assists in prediction of different diseases. These predictions are used to train a recommendation engine that uses Apriori and Fuzzy C Means (FCM) for suggesting corrective behavioral measures for a healthier lifestyle under real-time conditions. Due to these operations, the proposed model is able to improve behavior prediction accuracy by 16.4%, precision of prediction by 8.3%, AUC (area under the curve) of prediction by 9.5%, and accuracy of corrective behavior recommendation by 3.9% when compared with existing methods under similar evaluation conditions.

Synthetic 3D full-body skeletal motion from 2D paths using RNN with LSTM cells and linear networks.

Gait analysis has proven to be a key process in the functional assessment of people involving many fields, such as diagnosis of diseases or rehabilitation, and has increased in relevance lately. Gait analysis often requires gathering data, although this can be very expensive and time consuming. One of the main solutions applied in fields when data acquisition is a problem is augmentation of datasets with artificial data. There are two main approaches for doing that: simulation and synthetic data generation. In this article, we propose a parametrizable generative system of synthetic walking simplified human skeletons. For achieving that, a data gathering experiment with up to 26 individuals was conducted. The system consists of two artificial neural networks: a recurrent neural network for the generation of the movement and a multilayer perceptron for determining the size of the segments of the skeletons. The system has been evaluated through four processes: (i) an observational appraisal by researchers in gait analysis, (ii) a visual representation of the distribution of the generated data, (iii) a numerical analysis using the normalized cross-correlation coefficient, and (iv) an angular evaluation to check the kinematic validity of the data. The evaluation concluded that the system is able to generate realistic and accurate gait data. These results reveal a promising path for this research field, which can be further improved through increasing the variety of movements and the user sample.

Hybrid controller with neural network PID/FOPID operations for two-link rigid robot manipulator based on the zebra optimization algorithm.

The performance of the robotic manipulator is negatively impacted by outside disturbances and uncertain parameters. The system's variables are also highly coupled, complex, and nonlinear, indicating that it is a multi-input, multi-output system. Therefore, it is necessary to develop a controller that can control the variables in the system in order to handle these complications. This work proposes six control structures based on neural networks (NNs) with proportional integral derivative (PID) and fractional-order PID (FOPID) controllers to operate a 2-link rigid robot manipulator (2-LRRM) for trajectory tracking. These are named as set-point-weighted PID (W-PID), set-point weighted FOPID (W-FOPID), recurrent neural network (RNN)-like PID (RNNPID), RNN-like FOPID (RNN-FOPID), NN+PID, and NN+FOPID controllers. The zebra optimization algorithm (ZOA) was used to adjust the parameters of the proposed controllers while reducing the integral-time-square error (ITSE). A new objective function was proposed for tuning to generate controllers with minimal chattering in the control signal. After implementing the proposed controller designs, a comparative robustness study was conducted among these controllers by altering the initial conditions, disturbances, and model uncertainties. The simulation results demonstrate that the NN+FOPID controller has the best trajectory tracking performance with the minimum ITSE and best robustness against changes in the initial states, external disturbances, and parameter uncertainties compared to the other controllers.

Enhancement of cyber security in IoT based on ant colony optimized artificial neural adaptive Tensor flow.

The Internet of Things (IoT) is a network that connects various hardware, software, data storage, and applications. These interconnected devices provide services to businesses and can potentially serve as entry points for cyber-attacks. The privacy of IoT devices is increasingly vulnerable, particularly to threats like viruses and illegal software distribution lead to the theft of critical information. Ant Colony-Optimized Artificial Neural-Adaptive Tensorflow (ACO-ANT) technique is proposed to detect malicious software illicitly disseminated through the IoT. To emphasize the significance of each token in source duplicate data, the noise data undergoes processing using tokenization and weighted attribute techniques. Deep learning (DL) methods are then employed to identify source code duplication. Also the Multi-Objective Recurrent Neural Network (M-RNN) is used to identify suspicious activities within an IoT environment. The performance of proposed technique is examined using Loss, accuracy, F measure, precision to identify its efficiency. The experimental outcomes demonstrate that the proposed method ACO-ANT on Malimg dataset provides 12.35%, 14.75%, 11.84% higher precision and 10.95%, 15.78%, 13.89% higher f-measure compared to the existing methods. Further, leveraging block chain for malware detection is a promising direction for future research the fact that could enhance the security of IoT and identify malware threats.

Propofol anesthesia destabilizes neural dynamics across cortex.

Every day, hundreds of thousands of people undergo general anesthesia. One hypothesis is that anesthesia disrupts dynamic stability-the ability of the brain to balance excitability with the need to be stable and controllable. To test this hypothesis, we developed a method for quantifying changes in population-level dynamic stability in complex systems: delayed linear analysis for stability estimation (DeLASE). Propofol was used to transition animals between the awake state and anesthetized unconsciousness. DeLASE was applied to macaque cortex local field potentials (LFPs). We found that neural dynamics were more unstable in unconsciousness compared with the awake state. Cortical trajectories mirrored predictions from destabilized linear systems. We mimicked the effect of propofol in simulated neural networks by increasing inhibitory tone. This in turn destabilized the networks, as observed in the neural data. Our results suggest that anesthesia disrupts dynamical stability that is required for consciousness.

Infinite-dimensional reservoir computing.

Reservoir computing approximation and generalization bounds are proved for a new concept class of input/output systems that extends the so-called generalized Barron functionals to a dynamic context. This new class is characterized by the readouts with a certain integral representation built on infinite-dimensional state-space systems. It is shown that this class is very rich and possesses useful features and universal approximation properties. The reservoir architectures used for the approximation and estimation of elements in the new class are randomly generated echo state networks with either linear or ReLU activation functions. Their readouts are built using randomly generated neural networks in which only the output layer is trained (extreme learning machines or random feature neural networks). The results in the paper yield a recurrent neural network-based learning algorithm with provable convergence guarantees that do not suffer from the curse of dimensionality when learning input/output systems in the class of generalized Barron functionals and measuring the error in a mean-squared sense.

Accurate Neonatal Face Detection for Improved Pain Classification in the Challenging NICU Setting.

There is a tendency for object detection systems using off-the-shelf algorithms to fail when deployed in complex scenes. The present work describes a case for detecting facial expression in post-surgical neonates (newborns) as a modality for predicting and classifying severe pain in the Neonatal Intensive Care Unit (NICU). Our initial testing showed that both an off-the-shelf face detector and a machine learning algorithm trained on adult faces failed to detect facial expression of neonates in the NICU. We improved accuracy in this complex scene by training a state-of-the-art "You-Only-Look-Once" (YOLO) face detection model using the USF-MNPAD-I dataset of neonate faces. At run-time our trained YOLO model showed a difference of 8.6% mean Average Precision (mAP) and 21.2% Area under the ROC Curve (AUC) for automatic classification of neonatal pain compared with manual pain scoring by NICU nurses. Given the challenges, time and effort associated with collecting ground truth from the faces of post-surgical neonates, here we share the weights from training our YOLO model with these facial expression data. These weights can facilitate the further development of accurate strategies for detecting facial expression, which can be used to predict the time to pain onset in combination with other sensory modalities (body movements, crying frequency, vital signs). Reliable predictions of time to pain onset in turn create a therapeutic window of time wherein NICU nurses and providers can implement safe and effective strategies to mitigate severe pain in this vulnerable patient population.

De novo synthetic antimicrobial peptide design with a recurrent neural network.

Antibiotic resistance is recognized as an imminent and growing global health threat. New antimicrobial drugs are urgently needed due to the decreasing effectiveness of conventional small-molecule antibiotics. Antimicrobial peptides (AMPs), a class of host defense peptides, are emerging as promising candidates to address this need. The potential sequence space of amino acids is combinatorially vast, making it possible to extend the current arsenal of antimicrobial agents with a practically infinite number of new peptide-based candidates. However, mining naturally occurring AMPs, whether directly by wet lab screening methods or aided by bioinformatics prediction tools, has its theoretical limit regarding the number of samples or genomic/transcriptomic resources researchers have access to. Further, manually designing novel synthetic AMPs requires prior field knowledge, restricting its throughput. In silico sequence generation methods are gaining interest as a high-throughput solution to the problem. Here, we introduce AMPd-Up, a recurrent neural network based tool for de novo AMP design, and demonstrate its utility over existing methods. Validation of candidates designed by AMPd-Up through antimicrobial susceptibility testing revealed that 40 of the 58 generated sequences possessed antimicrobial activity against Escherichia coli and/or Staphylococcus aureus. These results illustrate that AMPd-Up can be used to design novel synthetic AMPs with potent activities.

BiRNN-DDI: A Drug-Drug Interaction Event Type Prediction Model Based on Bidirectional Recurrent Neural Network and Graph2Seq Representation.

Research on drug-drug interaction (DDI) prediction, particularly in identifying DDI event types, is crucial for understanding adverse drug reactions and drug combinations. This work introduces a Bidirectional Recurrent Neural Network model for DDI event type prediction (BiRNN-DDI), which simultaneously considers structural relationships and contextual information. Our BiRNN-DDI model constructs drug feature graphs to mine structural relationships. For contextual information, it transforms drug graphs into sequences and employs a two-channel structure, integrating BiRNN, to obtain contextual representations of drug-drug pairs. The model's effectiveness is demonstrated through comparisons with state-of-the-art models on two DDI event-type benchmarks. Extensive experimental results reveal that BiRNN-DDI surpasses other models in accuracy, AUPR, AUC, F1 score, Precision, and Recall metrics on both small and large datasets. Additionally, our model exhibits a lower parameter space, indicating more efficient learning of drug feature representations and prediction of potential DDI event types.

Gradient-free training of recurrent neural networks using random perturbations.

Recurrent neural networks (RNNs) hold immense potential for computations due to their Turing completeness and sequential processing capabilities, yet existing methods for their training encounter efficiency challenges. Backpropagation through time (BPTT), the prevailing method, extends the backpropagation (BP) algorithm by unrolling the RNN over time. However, this approach suffers from significant drawbacks, including the need to interleave forward and backward phases and store exact gradient information. Furthermore, BPTT has been shown to struggle to propagate gradient information for long sequences, leading to vanishing gradients. An alternative strategy to using gradient-based methods like BPTT involves stochastically approximating gradients through perturbation-based methods. This learning approach is exceptionally simple, necessitating only forward passes in the network and a global reinforcement signal as feedback. Despite its simplicity, the random nature of its updates typically leads to inefficient optimization, limiting its effectiveness in training neural networks. In this study, we present a new approach to perturbation-based learning in RNNs whose performance is competitive with BPTT, while maintaining the inherent advantages over gradient-based learning. To this end, we extend the recently introduced activity-based node perturbation (ANP) method to operate in the time domain, leading to more efficient learning and generalization. We subsequently conduct a range of experiments to validate our approach. Our results show similar performance, convergence time and scalability when compared to BPTT, strongly outperforming standard node perturbation and weight perturbation methods. These findings suggest that perturbation-based learning methods offer a versatile alternative to gradient-based methods for training RNNs which can be ideally suited for neuromorphic computing applications.

Research on prediction and evaluation algorithm of sports athletes performance based on neural network.

BACKGROUND: The Ultimate Fighting Championship (UFC) stands as a prominent global platform for professional mixed martial arts, captivating audiences worldwide. With its continuous growth and globalization efforts, UFC events have garnered significant attention and achieved commendable results. However, as the scale of development expands, the operational demands on UFC events intensify. At its core, UFC thrives on the exceptional performances of its athletes, which serve as the primary allure for audiences. OBJECTIVE: This study aims to enhance the allure of UFC matches and cultivate exceptional athletes by predicting athlete performance on the field. To achieve this, a recurrent neural network prediction model based on Bidirectional Long Short-Term Memory (BiLSTM) is proposed. The model seeks to leverage athlete portraits and characteristics for performance prediction. METHODS: The proposed methodology involves constructing athlete portraits and analyzing athlete characteristics to develop the prediction model. The BiLSTM-based recurrent neural network is utilized for its ability to capture temporal dependencies in sequential data. The model's performance is assessed through experimental analysis. RESULTS: Experimental results demonstrate that the athlete performance prediction model achieved an overall accuracy of 0.7524. Comparative analysis reveals that the proposed BiLSTM model outperforms traditional methods such as Linear Regression and Multilayer Perceptron (MLP), showcasing superior prediction accuracy. CONCLUSION: This study introduces a novel approach to predicting athlete performance in UFC matches using a BiLSTM-based recurrent neural network. By leveraging athlete portraits and characteristics, the proposed model offers improved accuracy compared to classical methods. Enhancing the predictive capabilities in UFC not only enriches the viewing experience but also contributes to the development of exceptional athletes in the sport.

DeepHLApan: A Deep Learning Approach for the Prediction of Peptide-HLA Binding and Immunogenicity.

Neoantigens are crucial in distinguishing cancer cells from normal ones and play a significant role in cancer immunotherapy. The field of bioinformatics prediction for tumor neoantigens has rapidly developed, focusing on the prediction of peptide-HLA binding affinity. In this chapter, we introduce a user-friendly tool named DeepHLApan, which utilizes deep learning techniques to predict neoantigens by considering both peptide-HLA binding affinity and immunogenicity. We provide the application of DeepHLApan, along with the source code, docker version, and web-server. These resources are freely available at https://github.com/zjupgx/deephlapan and http://pgx.zju.edu.cn/deephlapan/ .

Attention-Based Variational Autoencoder Models for Human-Human Interaction Recognition via Generation.

The remarkable human ability to predict others' intent during physical interactions develops at a very early age and is crucial for development. Intent prediction, defined as the simultaneous recognition and generation of human-human interactions, has many applications such as in assistive robotics, human-robot interaction, video and robotic surveillance, and autonomous driving. However, models for solving the problem are scarce. This paper proposes two attention-based agent models to predict the intent of interacting 3D skeletons by sampling them via a sequence of glimpses. The novelty of these agent models is that they are inherently multimodal, consisting of perceptual and proprioceptive pathways. The action (attention) is driven by the agent's generation error, and not by reinforcement. At each sampling instant, the agent completes the partially observed skeletal motion and infers the interaction class. It learns where and what to sample by minimizing the generation and classification errors. Extensive evaluation of our models is carried out on benchmark datasets and in comparison to a state-of-the-art model for intent prediction, which reveals that classification and generation accuracies of one of the proposed models are comparable to those of the state of the art even though our model contains fewer trainable parameters. The insights gained from our model designs can inform the development of efficient agents, the future of artificial intelligence (AI).

An Artificial Neural Network Predicts Gender Differences of Motor and Non-Motor Symptoms of Patients with Advanced Parkinson's Disease under Levodopa-Carbidopa Intestinal Gel.

Background and Objectives: Currently, no tool exists to predict clinical outcomes in patients with advanced Parkinson's disease (PD) under levodopa-carbidopa intestinal gel (LCIG) treatment. The aim of this study was to develop a novel deep neural network model to predict the clinical outcomes of patients with advanced PD after two years of LCIG therapy. Materials and Methods: This was a longitudinal, 24-month observational study of 59 patients with advanced PD in a multicenter registry under LCIG treatment from September 2019 to September 2021, including 43 movement disorder centers. The data set includes 649 measurements of patients, which make an irregular time series, and they are turned into regular time series during the preprocessing phase. Motor status was assessed with the Unified Parkinson's Disease Rating Scale (UPDRS) Parts III (off) and IV. The NMS was assessed by the NMS Questionnaire (NMSQ) and the Geriatric Depression Scale (GDS), the quality of life by PDQ-39, and severity by Hoehn and Yahr (HY). Multivariate linear regression, ARIMA, SARIMA, and Long Short-Term Memory-Recurrent NeuralNetwork (LSTM-RNN) models were used. Results: LCIG significantly improved dyskinesia duration and quality of life, with men experiencing a 19% and women a 10% greater improvement, respectively. Multivariate linear regression models showed that UPDRS-III decreased by 1.5 and 4.39 units per one-unit increase in the PDQ-39 and UPDRS-IV indexes, respectively. Although the ARIMA-(2,0,2) model is the best one with AIC criterion 101.8 and validation criteria MAE = 0.25, RMSE = 0.59, and RS = 0.49, it failed to predict PD patients' features over a long period of time. Among all the time series models, the LSTM-RNN model predicts these clinical characteristics with the highest accuracy (MAE = 0.057, RMSE = 0.079, RS = 0.0053, mean square error = 0.0069). Conclusions: The LSTM-RNN model predicts, with the highest accuracy, gender-dependent clinical outcomes in patients with advanced PD after two years of LCIG therapy.

Combined Control for a Piezoelectric Actuator Using a Feed-Forward Neural Network and Feedback Integral Fast Terminal Sliding Mode Control.

In recent years, there has been significant interest in incorporating micro-actuators into industrial environments; this interest is driven by advancements in fabrication methods. Piezoelectric actuators (PEAs) have emerged as vital components in various applications that require precise control and manipulation of mechanical systems. These actuators play a crucial role in the micro-positioning systems utilized in nanotechnology, microscopy, and semiconductor manufacturing; they enable extremely fine movements and adjustments and contribute to vibration control systems. More specifically, they are frequently used in precision positioning systems for optical components, mirrors, and lenses, and they enhance the accuracy of laser systems, telescopes, and image stabilization devices. Despite their numerous advantages, PEAs exhibit complex dynamics characterized by phenomena such as hysteresis, which can significantly impact accuracy and performance. The characterization of these non-linearities remains a challenge for PEA modeling. Recurrent artificial neural networks (ANNs) may simplify the modeling of the hysteresis dynamics for feed-forward compensation. To address these challenges, robust control strategies such as integral fast terminal sliding mode control (IFTSMC) have been proposed. Unlike traditional fast terminal sliding mode control methods, IFTSMC includes integral action to minimize steady-state errors, improving the tracking accuracy and disturbance rejection capabilities. However, accurate modeling of the non-linear dynamics of PEAs remains a challenge. In this study, we propose an ANN-based IFTSMC controller to address this issue and to enhance the precision and reliability of PEA positioning systems. We implement and validate the proposed controller in a real-time setup and compare its performance with that of a PID controller. The results obtained from real PEA experiments demonstrate the stability of the novel control structure, as corroborated by the theoretical analysis. Furthermore, experimental validation reveals a notable reduction in error compared to the PID controller.

Hybrid deep learning and optimized clustering mechanism for load balancing and fault tolerance in cloud computing.

Cloud services are one of the most quickly developing technologies. Furthermore, load balancing is recognized as a fundamental challenge for achieving energy efficiency. The primary function of load balancing is to deliver optimal services by releasing the load over multiple resources. Fault tolerance is being used to improve the reliability and accessibility of the network. In this paper, a hybrid Deep Learning-based load balancing algorithm is developed. Initially, tasks are allocated to all VMs in a round-robin method. Furthermore, the Deep Embedding Cluster (DEC) utilizes the Central Processing Unit (CPU), bandwidth, memory, processing elements, and frequency scaling factors while determining if a VM is overloaded or underloaded. The task performed on the overloaded VM is valued and the tasks accomplished on the overloaded VM are assigned to the underloaded VM for cloud load balancing. In addition, the Deep Q Recurrent Neural Network (DQRNN) is proposed to balance the load based on numerous factors such as supply, demand, capacity, load, resource utilization, and fault tolerance. Furthermore, the effectiveness of this model is assessed by load, capacity, resource consumption, and success rate, with ideal values of 0.147, 0.726, 0.527, and 0.895 are achieved.

An adaptive discretized RNN algorithm for posture collaboration motion control of constrained dual-arm robots.

Although there are many studies on repetitive motion control of robots, few schemes and algorithms involve posture collaboration motion control of constrained dual-arm robots in three-dimensional scenes, which can meet more complex work requirements. Therefore, this study establishes the minimum displacement repetitive motion control scheme for the left and right robotic arms separately. On the basis of this, the design mentality of the proposed dual-arm posture collaboration motion control (DAPCMC) scheme, which is combined with a new joint-limit conversion strategy, is described, and the scheme is transformed into a time-variant equation system (TVES) problem form subsequently. To address the TVES problem, a novel adaptive Taylor-type discretized recurrent neural network (ATT-DRNN) algorithm is devised, which fundamentally solves the problem of calculation accuracy which cannot be balanced well with the fast convergence speed. Then, stringent theoretical analysis confirms the dependability of the ATT-DRNN algorithm in terms of calculation precision and convergence rate. Finally, the effectiveness of the DAPCMC scheme and the excellent convergence competence of the ATT-DRNN algorithm is verified by a numerical simulation analysis and two control cases of dual-arm robots.

Sentiment analysis of the Hamas-Israel war on YouTube comments using deep learning.

Sentiment analysis aims to classify text based on the opinion or mentality expressed in a situation, which can be positive, negative, or neutral. Therefore, in the world, a lot of opinions are available on various social media sites, which must be gathered and analyzed to assess the general public's opinion. Finding and monitoring comments, as well as manually extracting the information contained in them, is a difficult task due to the vast diversity of ideas on YouTube. Identifying public opinion on war topics is crucial for offering insights to opposing sides based on popular opinion and emotions about the ongoing war. To address the gap, we build a model on YouTube comment sentiment analysis of the Hamas-Israel war to determine public opinion. In this study, we address the gaps by developing a deep learning-based approach for sentiment analysis. We have collected 24,360 comments from popular YouTube News Channels including BBC, WION, Aljazeera, and others about the Hamas-Israel War using YouTube API and Google spreadsheet and labeled them by linguistic experts into three classes: positive, negative, and neutral. Then, textual comments were preprocessed using natural language processing (NLP) techniques, and features were extracted using Word2vec, FastText, and GloVe. Moreover, we have used the SMOTE data balancing technique and used different data splits, but the 80/20 train-test split ratio has the highest accuracy. For classification model building, commonly used classification algorithms LSTM, Bi-LSTM, GRU, and Hybrid of CNN and Bi-LSTM were applied, and their performance is compared. As a result, the Hybrid of CNN and Bi-LSTM with Word2vec achieved the highest performance with 95.73% accuracy for comments classifications.

SJFO: Sail Jelly Fish Optimization enabled VM migration with DRNN-based prediction for load balancing in cloud computing.

The dynamic workload is evenly distributed among all nodes using balancing methods like hosts or VMs. Load Balancing as a Service (LBaaS) is another name for load balancing in the cloud. In this research work, the load is balanced by the application of Virtual Machine (VM) migration carried out by proposed Sail Jelly Fish Optimization (SJFO). The SJFO is formed by combining Sail Fish Optimizer (SFO) and Jellyfish Search (JS) optimizer. In the Cloud model, many Physical Machines (PMs) are present, where these PMs are comprised of many VMs. Each VM has many tasks, and these tasks depend on various parameters like Central Processing Unit (CPU), memory, Million Instructions per Second (MIPS), capacity, total number of processing entities, as well as bandwidth. Here, the load is predicted by Deep Recurrent Neural Network (DRNN) and this predicted load is compared with a threshold value, where VM migration is done based on predicted values. Furthermore, the performance of SJFO-VM is analysed using the metrics like capacity, load, and resource utilization. The proposed method shows better performance with a superior capacity of 0.598, an inferior load of 0.089, and an inferior resource utilization of 0.257.

A multi-resolution physics-informed recurrent neural network: formulation and application to musculoskeletal systems.

This work presents a multi-resolution physics-informed recurrent neural network (MR PI-RNN), for simultaneous prediction of musculoskeletal (MSK) motion and parameter identification of the MSK systems. The MSK application was selected as the model problem due to its challenging nature in mapping the high-frequency surface electromyography (sEMG) signals to the low-frequency body joint motion controlled by the MSK and muscle contraction dynamics. The proposed method utilizes the fast wavelet transform to decompose the mixed frequency input sEMG and output joint motion signals into nested multi-resolution signals. The prediction model is subsequently trained on coarser-scale input-output signals using a gated recurrent unit (GRU), and then the trained parameters are transferred to the next level of training with finer-scale signals. These training processes are repeated recursively under a transfer-learning fashion until the full-scale training (i.e., with unfiltered signals) is achieved, while satisfying the underlying dynamic equilibrium. Numerical examples on recorded subject data demonstrate the effectiveness of the proposed framework in generating a physics-informed forward-dynamics surrogate, which yields higher accuracy in motion predictions of elbow flexion-extension of an MSK system compared to the case with single-scale training. The framework is also capable of identifying muscle parameters that are physiologically consistent with the subject's kinematics data.