Fully-connected Neural Network Research Articles

A Comparative Analysis of Data Synthesis Techniques to Improve Classification Accuracy of Raman Spectroscopy Data.

Raman spectra are examples of high dimensional data that can often be limited in the number of samples. This is a primary concern when Deep Learning frameworks are developed for tasks such as chemical species identification, quantification, and diagnostics. Open-source data are difficult to obtain and often sparse; furthermore, the collecting and curating of new spectra require expertise and resources. Deep generative modeling utilizes Deep Learning architectures to approximate high dimensional distributions and aims to generate realistic synthetic data. The evaluation of the data and the performance of the deep models is usually conducted on a per-task basis and provides no indication of an increase to robustness, or generalization, on a wider scale. In this study, we compare the benefits and limitations of a standard statistical approach to data synthesis (weighted blending) with a popular deep generative model, the Variational Autoencoder. Two binary data sets are divided into 3-fold to simulate small, limited samples. Synthetic data distributions are created per fold using the two methods and then augmented into the training of two Deep Learning algorithms, a Convolutional Neural Network and a Fully-Connected Neural Network. The goal of this study is to observe the trends in learning as synthetic data are continually augmented to the training data in increasing batches. To determine the impact of each synthetic method, Principal Component Analysis and the discrete Fréchet distance are implemented to visualize and measure the distance between the source and synthetic distributions along with the Machine Learning metric balanced accuracy for evaluating performance on imbalanced data.

Generative modeling of living cells with SO(3)-equivariant implicit neural representations

Data-driven cell tracking and segmentation methods in biomedical imaging require diverse and information-rich training data. In cases where the number of training samples is limited, synthetic computer-generated data sets can be used to improve these methods. This requires the synthesis of cell shapes as well as corresponding microscopy images using generative models. To synthesize realistic living cell shapes, the shape representation used by the generative model should be able to accurately represent fine details and changes in topology, which are common in cells. These requirements are not met by 3D voxel masks, which are restricted in resolution, and polygon meshes, which do not easily model processes like cell growth and mitosis. In this work, we propose to represent living cell shapes as level sets of signed distance functions (SDFs) which are estimated by neural networks. We optimize a fully-connected neural network to provide an implicit representation of the SDF value at any point in a 3D+time domain, conditioned on a learned latent code that is disentangled from the rotation of the cell shape. We demonstrate the effectiveness of this approach on cells that exhibit rapid deformations (Platynereis dumerilii), cells that grow and divide (C. elegans), and cells that have growing and branching filopodial protrusions (A549 human lung carcinoma cells). A quantitative evaluation using shape features and Dice similarity coefficients of real and synthetic cell shapes shows that our model can generate topologically plausible complex cell shapes in 3D+time with high similarity to real living cell shapes. Finally, we show how microscopy images of living cells that correspond to our generated cell shapes can be synthesized using an image-to-image model.

Uniaxial constitutive model for fiber reinforced concrete: A physics-based data-driven framework

Fiber reinforced concrete (FRC) has improved strength and ductility, making it suitable for a wider range of engineering structures. The development of a constitutive model is crucial for analyzing mechanical behavior of these structures, while this still remains challenging as the complex composition of FRC makes it difficult to formulate explicit relationships among a range of critical material parameters. With the latest development of data and digital technologies, data-driven approaches have emerged as a powerful alternative that are capable of solving advanced and complex engineering problems. Developing data-driven methods based on objective data analysis and decision-making to predict mechanical behavior of engineering structures is an important direction of such development. In this paper, a uniaxial constitutive model of fiber-reinforced concrete is developed by a physics-based data-driven framework. The framework consists of three important parts, including construction of experimental database, parameters calibration of physical models, and implementation of neural network. From the experimental data collected from published literature, an experimental database is built first. A physical model is proposed by modifying the uniaxial constitute model for normal concrete, so that it is more convenient and straightforward to consider the effect of fibers on fiber reinforced concrete subjected to compression, tension and repeat loading conditions. The parameters of the model are calibrated against experimental data by the swarm intelligence optimization algorithms. With the calibrated parameters of the physical model, a Fully Connected Neural Network (FCNN) is trained to be used to predict the physical parameters of fiber reinforced concrete. By comparing with independent experimental data, the proposed fiber-reinforced concrete uniaxial constitutive model constructed using the physics-based data-driven framework can accurately predict the stress–strain relationships of a range of FRC, which suggests that the FRC material model can be used in the numerical simulation and design of FRC components and structures. In addition, the proposed model is applicable to multiple loading conditions.

A generative modeling framework for inferring families of biomechanical constitutive laws in data-sparse regimes

Quantifying biomechanical properties of the human vasculature could deepen our understanding of cardiovascular diseases. Standard nonlinear regression in constitutive modeling requires considerable high-quality data and an explicit form of the constitutive model as prior knowledge. By contrast, we propose a novel approach that combines generative deep learning with Bayesian inference to efficiently infer families of constitutive relationships in data-sparse regimes. Inspired by the concept of functional priors, we develop a generative adversarial network (GAN) that incorporates a neural operator as the generator and a fully-connected neural network as the discriminator. The generator takes a vector of noise conditioned on measurement data as input and yields the predicted constitutive relationship, which is scrutinized by the discriminator in the following step. We demonstrate that this framework can accurately estimate means and standard deviations of the constitutive relationships of the murine aorta using data collected either from model-generated synthetic data or ex vivo experiments for mice with genetic deficiencies. In addition, the framework learns priors of constitutive models without explicitly knowing their functional form, providing a new model-agnostic approach to learning hidden constitutive behaviors from data.

Learning Mappings from Iced Airfoils to Aerodynamic Coefficients Using a Deep Operator Network

This study abstracted the prediction of the aerodynamic coefficients of an iced airfoil as a mapping from the iced airfoil space to the aerodynamic coefficient space. Thus, a deep network called Airfoils2AeroNet that can learn this mapping was established based on Deep Operator Network (DeepONet). The deep network consists of a branch network for encoding the iced airfoil images and a trunk network that learns a nonlinear mapping from the one-dimensional aerodynamic coefficient function input to p-dimensional outputs. The branch network consists of deep convolutional neural networks (CNNs), and the trunk network consists of fully connected neural networks (FNNs). Then the network was trained and tested on iced airfoils based on NACA 0012 airfoils. Comparing the prediction results of Airfoils2AeroNet with those of the conventional direct CNN network, the network proposed in this paper has a strong advantage for generalization. Unlike the traditional CNN, which can only predict the aerodynamic coefficients at fixed flow conditions consistent with the training data, the network can flexibly predict the aerodynamic coefficients at different flow conditions. Finally, the influence of the structure of the branch network and trunk network on the prediction results was analyzed.

Correlation between figure of merit function and deviation in the assessment of kinetic parameters in realistic TLD-100 behavior using fully-connected neural networks

In this study, a realistic model was employed to research the LiF:Mg:Ti behavior (commercially known as TLD-100), a widely employed as thermoluminescent (TL) material in dosimetry services. The research was focused on exploring the association between the Figure of Merit (FOM) function and deviations from the true kinetic parameter values, that were obtained through fitting experimental measures. FOM function is the most used magnitude to measure the quality of a fit achieved between a mathematical model and experimental data. Indeed, FOM function is particularly relevant when the TL glow curve deconvolution method is implemented to obtain the kinetic parameters: Activation Energy E (eV), Frequency Factor s (s−1), value of kinetic order b, Temperature at which the intensity maximum TMax(oC), Intensity Maximum IMax (a.u.) and Distribution Width σ (eV) of every peak. FOM function is subject to degeneracy under certain conditions, this means that different sets of kinetic parameters can lead to similar FOM values. Furthermore, as the obtained kinetic parameters are deemed acceptable when the FOM value is below 5%, then the physical degeneracy is greater than the mathematical ones. This contribution provides a rigorous analysis of how the FOM values can be impacted by deviations between experimental and estimated data of kinetic parameters from the analysis process. This studio was carried out using a mathematical method based on Artificial Intelligence with a high capability to manage a large volume of data. Another advantage of this method is that the analysis to be extended to a wide range of experimental data.

A heuristic tomato-bunch harvest manipulator path planning method based on a 3D-CNN-based position posture map and rapidly-exploring random tree

The environment of tomato bunches is complex, and the fruit volume is relatively large, so manipulator picking-harvest motion planning should consider not only how to pick, but also how to avoid obstacles after picking tomato bunches and harvesting them from the complex environment, which puts forward strict requirements for picking-harvest motion performance. However, existing sampling-based algorithms still have problems of blind expansion and low efficiency. To solve these problems, a heuristic tomato-bunch harvest manipulator path planning method based on a 3D-CNN (convolution neural network)-based position posture map (PPM) and rapidly-exploring random tree (RRT) is proposed in this paper. This method comprehensively considers the sampling effectiveness of the whole process of picking and harvesting, search speed, path cost and manipulability of the manipulator, combining a sampling-based path planning algorithm and deep learning framework. The method utilizes CNN and FCN(Fully Connected Neural Networks) for generating a heuristic cost map where each voxel has a cost-to-go value toward the target and an optimal target posture value to guide the expansion tree. In addition, the sampling space of obstacle avoidance planning is reduced by a spatial partition method, and the difficulty caused by the increase in the volume of the end-effector clamping fruit is solved by prioritizing the planning of the fruit harvest path and then planning the picking path. On the basis of the above research, a large number of harvest experiments have verified the effectiveness of the proposed algorithm. Research and experimental results show that, based on the PPM-RRT algorithm, the picking-harvest time of a single tomato bunch is 13 s, and the success rate is close to 100%. Compared with state-of-the-art sampling algorithms, the planning time is reduced by more than 70%, the path length is reduced by more than 24%, and the manipulability of the manipulator is increased by 38.9%. The proposed path planning method is effective in complex environments.

A data-driven evaluating method on the defrosting effect of the air source heat pump system in Beijing

The current demand-based defrosting initiation strategies of air source heat pumps (ASHPs) faces difficulty in achieving a balance between cost and robustness. Therefore, an efficient and innovative demand-based defrosting initiation strategy is proposed in this paper, based on the learned the degradation of heating capacity (DHC) method using data-driven model. Firstly, the DHC method is developed using the initially installed sensors without additional sensors to identify the frosty state. Subsequently, the fully connected neural network (FNN) model is established for predicting frosty state using field measured data from the ASHP system. Finally, the effectiveness of the defrosting initiation strategy is further validated through practical testing. The results demonstrate that the DHC method provides a reliable database for training the FNN model, resulting in an impressive accuracy of 91.43% for predicting the frosty state in the testing set. By adopting the innovative demand-based defrosting initiation strategies, the defrosting frequency, heating loss, and power consumption are respectively decreased by 66.3%, 2.1%, and 6.0% with the SCOP enhanced by 8.6% during a heating season. The promising results highlight that the proposed strategy effectively reduces costs while ensuring the efficiency of demand-based control.

Learning ability of interpolating deep convolutional neural networks

It is frequently observed that overparameterized neural networks generalize well. Regarding such phenomena, existing theoretical work mainly devotes to linear settings or fully-connected neural networks. This paper studies the learning ability of an important family of deep neural networks, deep convolutional neural networks (DCNNs), under both underparameterized and overparameterized settings. We establish the first learning rates of underparameterized DCNNs without parameter or function variable structure restrictions presented in the literature. We also show that by adding well-defined layers to a non-interpolating DCNN, we can obtain some interpolating DCNNs that maintain the good learning rates of the non-interpolating DCNN. This result is achieved by a novel network deepening scheme designed for DCNNs. Our work provides theoretical verification of how overfitted DCNNs generalize well.

Deep multi-feature fusion residual network for oral squamous cell carcinoma classification and its intelligent system using Raman spectroscopy

Using fiber optic Raman spectroscopy and deep neural networks, we develop an intelligent system which will be used to assist surgeons accurately and efficiently to identify oral squamous cell carcinomas (OSCC). This system is able to classify 6 types of oral tissues. To achieve this goal, a novel classification framework called deep multi-feature fusion residual network (DMFF-ResNet) is proposed. This model is based on 16,200 Raman spectral data, obtained from the normal oral tissues and the OSCC of 90 patients through the surgical resection. Firstly, the 1-dimensional RestNet50 is taken as its backbone network. Then, the output spectral features of last three blocks are extracted from backbone network for feature fusion, which is expected to learn more spatial representations and have more discriminative power. Lastly, the derived spectral features are sent into a fully-connected neural network for performing the multiclassification task. Experimental results show that the proposed model achieves a competitive classification performance compared with state-of-the-art classifiers, and its accuracy, precision, and sensitivity reach 93.28%, 93.53%, and 93.13%, respectively. Further, the proposed framework is deployed on an edge computing device to form a prototype intelligent system for OSCC detection. To validate this system, we perform an offline test experiment in another 20 patients which demonstrates the developed intelligent system can successfully discriminate OSCC and normal oral tissues, with accuracy, precision, and recall of 92.78%, 92.33%, and 92.57%, respectively. The code was available at https://github.com/ISCLab-Bistu/retinanet-OSCC.

Deep learning-based design model for suction caissons on clay

Predicting the non-linear loading response is the key to the design of suction caissons. This paper presents a systematic study to explore the applicability of deep learning techniques in foundation design. Firstly, a series of three-dimensional finite element simulations was performed, covering a wide range of embedment ratios and different loading directions, to provide training data for the deep neural network (DNN) model. Then, hyper-parameter tuning was performed and it is found that the basic Fully-Connected (FC) neural network model is sufficient to capture the non-linear response of suction caissons with excellent accuracy and robustness. Furthermore, the optimized FC neural network model was also successfully applied to a database of suction caissons in sand, demonstrating its broad applicability. By comparing three typical DNNs, i.e., FC, Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM), it was observed that the FC neural network model excels over others in terms of simplicity, efficiency and accuracy. More importantly, by looking into the model’s generalization performance, the FC neural network model can also identify the change in foundation failure mechanisms. This study demonstrates the DNN’s powerful mapping ability and its potential for future use in offshore foundation design.

Physical-inforced artificial intelligent model for prediction of water-hammar velocity

The water hammer wave diagnostic method is a new promising fracture diagnostic method that can determine the fracture location underground. The key to this technology is the accurate calculation of the water-hammer pressure wave velocity (WHPWV). However, it is difficult to measure the pump shut-in WHPWV at hydraulic fracturing field. Furthermore, The commonly used WHPWV analytical prediction models are not able to comprehensively consider relevant factors, resulting in low prediction accuracy in construction site. Fully connected neural network (FCNN) as one of the artificial neural network (ANN), can discover the complex function relationship between lots of factors while training data is sufficient. But there is few data related to water-hammer in field of reservoir stimulation, which significantly limits the performance of FCNN. The existing study pointed out the prior imformation is beneficial to improve performance of ANN. In this article, a physical-inforced neural network is built to predict the WHPWV, in which the analytical model is used as prior information to construct the hidden layer called substructure mimic layer and full-structure mimic layer. Comparing with traditional FCNN, neural network we built have lower data cost, more accurate prediction and more stable performance, because the possible data mapping relationships are considered in advance. Furthermore, the mapping relationships implied by the model structure can be adapted to the actual situation. Due to the limited field data, a theoretical dataset was constructed to validate the model performance under different training environment.

A Battery Capacity Estimation Framework Combining Hybrid Deep Neural Network and Regional Capacity Calculation Based on Real-World Operating Data

Efficient battery capacity estimation is of utmost importance for safe and reliable operations of electric vehicles (EVs). This paper proposes a battery capacity estimation framework based on real-world EV operating data collected from forty electric buses of the same model operating in two cities. First, a reference capacity calculation method is presented by combining the Coulomb counting method with the incremental capacity analysis method. Then the impacts of temperature, current, and State-of-Charge on battery degradation are quantitatively analyzed. Using the historical probability distributions as battery health features, a hybrid deep neural network model that combines a convolutional neural network with a fully-connected neural network is proposed for battery capacity estimation. The validation results show that the proposed model outperforms the state-of-the-art methods and reaches a mean absolute percentage error of 2.79 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> , while maintaining low computational cost.

Compromise privacy in large-batch Federated Learning via model poisoning

Federated Learning (FL) is a distributed learning paradigm in which clients collaboratively train a global model with shared gradients while preserving the privacy of local data. Recent researches found that an adversary can reveal private local data with gradient inversion attacks. However, setting a large batchsize in local training can defense against these attacks effectively by confusing gradients computed on private data. Although some advanced methods have been proposed to improve the performance of gradient inversion attacks on the large-batch training, they are limited to specific model architectures (e.g., fully connected neural networks (FCNNs) with ReLU layers). To address these problems, we propose a novel gradient inversion attack to compromise privacy in large-batch FL via model poisoning. We poison clients' models with malicious parameters, which are constructed purposely to mitigate the confusion of aggregated gradients. For FCNNs, the private data can be perfectly recovered by analyzing gradients of the first fully connected (FC) layer. For convolutional neural networks (CNNs), we extend our proposed method as a hybrid approach, consisting of the analytic method and the optimization-based method. We first recover the feature maps after the convolutional layers and then reconstruct private data by minimizing the loss of data-wise feature map matching. We demonstrate the effectiveness of our proposed method on four datasets and show that it outperforms previous methods for large-batch FL (e.g., 64, 128, 256) and models with different activation layers (e.g., ReLU, Sigmoid and Tanh).

Discrimination of piled-up neutron-gamma pulses using charge comparison method and neural network for CLYC detectors

To realize neutron-gamma discrimination of piled-up pulses in high counting rate circumstances, the discrimination performance of the charge comparison method and three neural network models to actual piled-up pulses was studied in this paper. The neural network models, including a residual neural network (ResNet), a convolutional neural network (CNN) and a fully connected neural network (FCNN), were trained and tested by seven kinds of actual experimental data sets, which included the background signals (bg), non-piled-up neutron (n) and gamma (g) pulses as well as piled-up neutron + neutron (n + n), neutron + gamma (n + g), gamma + neutron (g + n), and gamma + gamma (g + g) pulses. The labels of the seven data sets were provided after discriminating by the charge comparison method. The results showed that the charge comparison method could be applied to discriminate piled-up neutron-gamma pulses with the Figure of Merit (FoM) value of 1.0. The integration length has been optimized to be different for piled-up and non-piled-up pulses, but follow the same criterion for neutron and gamma discrimination. The FoM values are worse than that of non-piled-up signals because the baseline fluctuation is very large under high counting rate conditions, which gives lots of interference to pulse shape discrimination. The ResNet model has the highest total prediction accuracy (93.85%) for the seven signal types, and through the comparison and analysis of the inconsistent events discriminated by the charge comparison method and the ResNet model, it is concluded that the discrimination result of the ResNet model is more accurate. These results indicate that both the charge comparison method and residual neural network can be used for complicated n/γ discrimination under high counting rate conditions, and the residual neural network works better.

GOCI-II geostationary satellite hourly aerosol optical depth obtained by data-driven methods: Validation and comparison

The Geostationary Ocean Color Imager II (GOCI-II) deployed on board GEO-KOMPSAT 2B (GK2B) satellite can observe the earth with high frequency, which is of great significance to the dynamic monitoring of regional aerosol optical depth (AOD). However, the lack of short-wave infrared bands makes the classical methods based on radiative transfer analysis difficult to achieve retrieval of aerosols. The data-driven methods based on learning from massive samples can achieve high-accuracy retrieval of aerosols on the local, regional and global scale. However, the samples required for the data-driven method are difficult to obtain due to the short operation time of the GK2B satellite and insufficient ground-based measurements. Aiming to solve such a problem, the Simplified and Robust Surface Reflectance Estimation and Simplified Aerosol Retrieval Algorithm (SEMARA) were aggregated to be applied in the study. The SEMARA method using ground-based observation can retrieve local AOD with high reliability and accuracy, which can be used to construct high-quality and high-quantity samples for training with higher efficiency compared with traditional sample construct methods. Four data-driven methods: Fully Connected Neural Network (FCNN), Random Forest (RF), Support Vector Machine (SVM) and eXtreme Gradient Boosting (XGBoost), were chosen to retrieve the AOD over South Korea in 2021 and 2022 based on GOCI-II. The results were validated and compared using ground-based measurements from the Aerosol Robotic Network (AENRONET) sites and show that the FCNN method has high accuracy and stability, followed by, the RF and XGBoost methods, while the SVM results significantly overestimate the spatial distribution of the AOD.

A fully connected neural Network-based voltage compensation method for accurate reconstruction in brain electrical impedance tomography

Acute cerebral ischemia is commonly accompanied by cerebral edema. With brain electrical impedance tomography (EIT), it is possible to monitor conductivity variation during dehydration treatment of the cerebral edema. However, the scalp is also dehydrated when dehydrating brain tissue which exerts a large impact on boundary measurement. Consequently, reconstruction quality of cerebral edema is seriously affected. To reduce the influence of scalp dehydration, a voltage data compensation method based on a fully connected neural network (FCNN) is proposed in this work. With this method, the voltage measured when brain tissue and scalp are simultaneously dehydrated can be compensated. To evaluate the performance of the proposed method, reconstruction of simulated cerebral ischemia in a three-layer head model is conducted. Images reconstructed with the compensated voltage data are compared with the results when there is no compensation. Besides, quantitative comparison is performed. The results indicate that it is almost impossible to identify the simulated cerebral ischemia without voltage compensation. In contrast, cerebral ischemia can be much more accurately reconstructed when the measured voltage is compensated with the proposed method.

Imaging and Characterization of Cement Annulus and Bonding Interfaces in Cased-wells with Fully Connected Neural Network

Well integrity evaluation is of great significance to the production life and productivity of oil and gas wells, as well as the protection of reservoir layers. However, traditional sonic logging technology has been unable to meet the current requirement for evaluating unconventional oil and gas resources. For instance, it falls short when it comes to detecting cementing quality between the cement and formation, as well as achieving a quantitative assessment of annulus. In addition, factors such as mud invasion and wellbore damage have the potential to compromise the integrity of the cement sheath and lead to radial alteration in the medium in the wellbore, which can affect cementing quality. To address these issues, a novel approach based on deep learning is proposed to reconstruct slowness models from wellbore data using a fully connected neural network (FCNN) to evaluate the cementing quality. This interpretation workflow requires forward modeling with the three-dimensional dyadic Green’s function to explore borehole acoustic signals and thus generate training and test datasets. Afterward, the spatial feature maps are learned to reconstruct the slowness models by utilizing all the wellbore data. Additionally, the Real-ESRGAN method is applied to borehole imaging to further improve the above inversion. The inversion results regarding the slowness value, wellbore structure, and interface are more consistent with the true model. Finally, the image pixels are optimized using the thresholding method, resulting in an image of superior quality in terms of sharpness and detail compared to previous results. This research reveals that the FCNN-based method can effectively invert the radial profile of slowness around the wellbore from monopole sonic logging data, and the Real-ESRGAN method can further enhance the imaging results to better approximate the true models. This method has also been applied to realistic borehole scenario models, demonstrating its feasibility.

Innovative design of metamaterial perfect absorbers via residual fully connected neural network modeling

In recent years, significant progress has been made in the research of metamaterial perfect absorbers for on-demand design based on deep learning methods, and how to improve the generalization ability and prediction accuracy of network models in inverse design has been a hot topic of recent research. In this paper, we propose a fully connected neural network (RFC-NN) based on the residual principle applied to the inverse design of perfect absorbers. We design the RFC-NN model by applying the residual network structure commonly used on convolutional neural networks to a fully connected neural network, and the regression coefficients of the predicted structural parameters and the corresponding absorption spectra in the inverse design are 0.988 and 0.923, respectively, which have better prediction performance and generalization ability than the ordinary fully connected neural network (FC-NN), tandem neural network (TNN), and one-dimensional residual convolutional neural network (1D-Resnet). Then, we successfully designed the metamaterial perfect absorber with an absorption bandwidth (the spectral range where the absorptivity is greater than 90%) of 1935 nm using the trained RFC-NN model, which corresponds to an average absorptivity of 93.70%. Meanwhile, we use the analysis method of the Pearson correlation coefficient to improve the accuracy of the RFC-NN model after inverse design. Our proposed design method proves to be very effective and can also be applied to the design of other types of functional nanophotonic devices.

Effect of Gait Speed on Trajectory Prediction Using Deep Learning Models for Exoskeleton Applications

Gait speed is an important biomechanical determinant of gait patterns, with joint kinematics being influenced by it. This study aims to explore the effectiveness of fully connected neural networks (FCNNs), with a potential application for exoskeleton control, in predicting gait trajectories at varying speeds (specifically, hip, knee, and ankle angles in the sagittal plane for both limbs). This study is based on a dataset from 22 healthy adults walking at 28 different speeds ranging from 0.5 to 1.85 m/s. Four FCNNs (a generalised-speed model, a low-speed model, a high-speed model, and a low-high-speed model) are evaluated to assess their predictive performance on gait speeds included in the training speed range and on speeds that have been excluded from it. The evaluation involves short-term (one-step-ahead) predictions and long-term (200-time-step) recursive predictions. The results show that the performance of the low- and high-speed models, measured using the mean absolute error (MAE), decreased by approximately 43.7% to 90.7% when tested on the excluded speeds. Meanwhile, when tested on the excluded medium speeds, the performance of the low-high-speed model improved by 2.8% for short-term predictions and 9.8% for long-term predictions. These findings suggest that FCNNs are capable of interpolating to speeds within the maximum and minimum training speed ranges, even if not explicitly trained on those speeds. However, their predictive performance decreases for gaits at speeds beyond or below the maximum and minimum training speed ranges.