Performance Of Deep Neural Networks Research Articles

Using ScrutinAI for visual inspection of DNN performance in a medical use case

Our Visual Analytics (VA) tool ScrutinAI supports human analysts to investigate interactively model performance and data sets. Model performance depends on labeling quality to a large extent. In particular in medical settings, generation of high quality labels requires in depth expert knowledge and is very costly. Often, data sets are labeled by collecting opinions of groups of experts. We use our VA tool to analyze the influence of label variations between different experts on the model performance. ScrutinAI facilitates to perform a root cause analysis that distinguishes weaknesses of deep neural network (DNN) models caused by varying or missing labeling quality from true weaknesses. We scrutinize the overall detection of intracranial hemorrhages and the more subtle differentiation between subtypes in a publicly available data set.

Fragility, robustness and antifragility in deep learning

We propose a systematic analysis of deep neural networks (DNNs) based on a signal processing technique for network parameter removal, in the form of synaptic filters that identifies the fragility, robustness and antifragility characteristics of DNN parameters. Our proposed analysis investigates if the DNN performance is impacted negatively, invariantly, or positively on both clean and adversarially perturbed test datasets when the DNN undergoes synaptic filtering. We define three filtering scores for quantifying the fragility, robustness and antifragility characteristics of DNN parameters based on the performances for (i) clean dataset, (ii) adversarial dataset, and (iii) the difference in performances of clean and adversarial datasets. We validate the proposed systematic analysis on ResNet-18, ResNet-50, SqueezeNet-v1.1 and ShuffleNet V2 x1.0 network architectures for MNIST, CIFAR10 and Tiny ImageNet datasets. The filtering scores, for a given network architecture, identify network parameters that are invariant in characteristics across different datasets over learning epochs. Vice-versa, for a given dataset, the filtering scores identify the parameters that are invariant in characteristics across different network architectures. We show that our synaptic filtering method improves the test accuracy of ResNet and ShuffleNet models on adversarial dataset when only the robust and antifragile parameters are selectively retrained at any given epoch, thus demonstrating applications of the proposed strategy in improving model robustness.

Enhancement of Deep Neural Network Recognition on MPSoC with Single Event Upset.

This paper introduces a new finding regarding single event upsets (SEUs) in configuration memory, and their potential impact on enhancing the performance of deep neural networks (DNNs) on the multiprocessor system on chip (MPSoC) platform. Traditionally, SEUs are considered to have negative effects on electronic systems or designs, but the current study demonstrates that they can also have positive contributions to the DNN on the MPSoC. The assertion that SEUs can have positive contributions to electronic system design was supported by conducting fault injections through dynamic reconfiguration on DNNs implemented on a 16nm FinFET technology Zynq UltraScale+ MPSoC. The results of the current study were highly significant, indicating that an SEU in configuration memory could result in an impressive 8.72% enhancement in DNN recognition on the MPSoC. One possible cause is that SEU in the configuration memory leads to slight changes in weight or bias values, resulting in improved activation levels of neurons and enhanced final recognition accuracy. This discovery offers a flexible and effective solution for boosting DNN performance on the MPSoC platform.

Modelling tree biomass using direct and additive methods with point cloud deep learning in a temperate mixed forest

Airborne laser scanning (ALS) data has been widely used for total aboveground tree biomass (AGB) modelling, however, there is less research focusing on estimating specific tree biomass components (wood, branches, bark, and foliage). Knowledge about these biomass components is essential for carbon accounting, understanding forest nutrient cycling, and other applications. In this study, we compare additive AGB estimation (sum of estimated components) with direct AGB estimation using deep neural network (DNN) and random forest (RF) models. We utilise two point cloud DNNs: point-based Dynamic Graph Convolutional Neural Network (DGCNN) and Octree-based Convolutional Neural Network (OCNN). DNN and RF models were trained using a dataset comprised of 2336 sample plots from a mixed temperate forest in New Brunswick, Canada. Results indicate that additive AGB models perform similarly to direct models in terms of coefficient of determination (R2) and root-mean square error (RMSE), and reduced the mean absolute percentage error (MAPE) by 22% on average. Compared to RF, the DNNs provided a small improvement in performance, with OCNN explaining 5% more variation in the data (R2 = 0.76) and reducing MAPE by 20% on average. Overall, this study showcases the effectiveness of additive tree AGB models and highlights the potential of DNNs for enhanced AGB estimation. To further improve DNN performance, we recommend using larger training datasets, implementing hyperparameter optimization, and incorporating additional data such as multispectral imagery.

Comprehensive analysis of the heterogeneous computing performance of DNNs under typical frameworks on cloud and edge computing platforms

In this paper, nearly 40 commonly used deep neural network(DNN) models are selected, and their cross-platform and cross-inference frameworks are deeply analysed. The main metrics of accuracy, the total number of model parameters, the computational complexity, the accuracy density, the inference time, the memory consumption and other related parameters are used to measure their performance. The heterogeneous computing experiment is implemented on both the Google Colab cloud computing platform and the Jetson Nano embedded edge computing platform. The obtained performance is compared with that of two previous computing platforms: a workstation equipped with an NVIDIA Titan X Pascal and an embedded system based on an NVIDIA Jetson TX1 board. In addition, on the Jetson Nano embedded edge computing platform, different inference frameworks are investigated to evaluate the inference efficiency of the DNN models. Regression models are established to characterize the variation in the computing performance of different DNN classification algorithms so that the inference results of unknown models can be estimated. ANOVA methods are proposed to quantify the differences between models. The experimental results have important guiding significance for the better selection, deployment and application of DNN models in practice. Codes are available at this https URL https://github.com/Foreverzfy/Model-Test.

Beamforming and data detection in intelligent reflecting surfaces‐assisted communication using deep learning

SummaryIn this article, we propose two deep neural network (DNN) architectures for an intelligent reflecting surface (IRS)‐assisted network with a single transmitter and multiple receivers. The first DNN is designed to obtain an optimal beamforming weight vector that maximizes the achievable sum rate over the IRS‐aided network. The beamforming vector is designed based on a pilot‐based channel estimation procedure, which requires only a small fraction of the available number of reflecting surfaces on the IRS to be active. The obtained vector is fed to the second DNN, which is designed for detecting information symbols from the transmitter. Our simulation study shows that the achievable rate obtained using the designed beamforming vector approaches the maximum achievable rate quickly, as the number of active elements increases. The impact of the number of active elements and the beamforming vector on the performance of the detector is also studied, in terms of the bit error rate (BER). We show that the BER performance of the second DNN is close to that of the optimal maximum likelihood detector. Further, we study the effect of the channel estimation errors on the performance of the DNNs. Moreover, we study the effect of the hyperparameters of the architectures on the training time and performance, in terms of root mean squared error and accuracy.

Deep predictions and transfer learning for simulation-driven structural health monitoring based on guided waves

Simulation-driven structural health monitoring (SHM) is crucial for enhancing the safety and reliability of structures. Its primary aim is to utilize numerical simulations to design efficient SHM systems, reducing costs while maintaining effectiveness. However, challenges arise due to disparities between simulations and physical measurements, stemming from uncertainties, noise, and simplifications. These challenges also extend to SHM methods using machine learning, particularly deep neural networks (DNNs).This paper bridges the gap between DNNs trained on simulated data and real-world experiments by investigating three transfer learning (TL) techniques: feature augmentation, direct fine-tuning, and boosting-based fine-tuning. The focus is on guided wave damage interaction coefficients (WDICs) as damage-sensitive features, modeled by DNNs for damage identification in thin plates, commonly found in aerospace structures like aircraft fuselages. This classification task compares measured and predicted WDICs in a common guided wave SHM configuration with one actuator and three sensors. Training and extensive testing datasets are generated using advanced finite element simulations and laser Doppler vibrometer experiments.The study demonstrates that TL techniques significantly enhance DNN prediction accuracy and improve damage identification performance for untrained scenarios compared to purely simulation-based DNNs. Among the TL approaches explored, feature augmentation with appropriate activation functions and boosting-based fine-tuning yield the most promising results. Using four out of twelve TL datasets, feature augmentation reduces the simulation-based mean absolute percentage error from 46.80% to 35.98% and increases coefficient of determination R2 values from 0.02 to 0.72. With six TL datasets and a boosting-based fine tuning, the mean absolute percentage error decreases from 47.42% to 35.20%, maintaining similar R2 values. While TL substantially enhances prediction accuracy, its impact on damage identification is slightly less pronounced, with improvements in relative classification accuracy by approximately 5% compared to simulation-based DNNs.These findings highlight the efficacy of TL techniques for damage identification in thin plates using WDICs and DNNs, emphasizing the influence of TL training and DNN model variations. Furthermore, they showcase the potential of TL methods to enhance DNN performance in future SHM applications, particularly emphasizing the promise of feature augmentation and boosting.

Adversarial attack and training for deep neural network based power quality disturbance classification

Power quality disturbance (PQD) can significantly affect the normal operation of the power system. Deep neural network (DNN) can classify PQD with extremely high accuracy. However, adversarial attacks can drastically degrade the performance of DNN, which raises security issues for DNN-based PQD classification. At present, there are few researches on adversarial attacks and their defenses in PQD classification. In this study, we first adopt convolutional neural network-long short-term memory (CNN-LSTM) to classify PQD signals. Then we propose an adversarial attack algorithm for PQD classification, i.e., momentum iterative-fast gradient sign method (MI-FGSM). MI-FGSM generates adversarial perturbations along the gradient direction of the loss function by incorporating momentum during iterations. Finally, we propose a defense algorithm to defend against adversarial attacks, i.e., iterative adversarial training (IAT). IAT can improve the robustness of the classification model by making it learn to minimize the worst loss caused by multi-class adversarial perturbation strengths. The experimental results demonstrate that compared with the fast gradient sign method, the MI-FGSM produces smaller perturbations and can further reduce the classification accuracy of CNN-LSTM. In addition, compared with the most advanced adversarial training defense algorithm in PQD classification, IAT can effectively enhance the robustness of the CNN-LSTM against adversarial attacks.

Towards Better Generalization of Deep Neural Networks via Non-Typicality Sampling Scheme.

Improving the generalization performance of deep neural networks (DNNs) trained by minibatch stochastic gradient descent (SGD) has raised lots of concerns from deep learning practitioners. The standard simple random sampling (SRS) scheme used in minibatch SGD treats all training samples equally in gradient estimation. In this article, we study a new data selection method based on the intrinsic property of the training set to help DNNs have better generalization performance. Our theoretical analysis suggests that this new sampling scheme, called the nontypicality sampling scheme, boosts the generalization performance of DNNs through biasing the solution toward wider minima, under certain assumptions. We confirm our findings experimentally and show that more variants of minibatch SGD can also benefit from the new sampling scheme. Finally, we discuss an extension of the nontypicality sampling scheme that holds promise to enhance both generalization performance and convergence speed of minibatch SGD.

Improved Giza pyramids construction algorithm for Modify the deep neural network-based method for energy demand forecasting

Accurate prediction of energy demand is crucial for improving services, reducing costs, and optimizing operations in energy systems. Deep neural networks (DNNs) have emerged as a popular method for energy demand forecasting. However, the performance of DNNs can be affected by data quality and hyperparameter selection. To address these concerns, this study proposes a novel energy demand forecasting technique that combines DNNs with an enhanced Giza pyramid construction methodology. The aim of this study is to provide a more reliable and effective approach for forecasting energy demand. The DNNs are employed to capture the complex relationships between input and output variables, while the Giza pyramids algorithm is utilized to optimal selection of hyperparameters of the network. Real-world energy demand data is used to evaluate the proposed approach, comparing it with state-of-the-art baseline models. The research methodology involves assessing the suggested approach using real-world energy demand information and conducting a comparative analysis with cutting-edge baseline models, including modified BP neural network (MBPNN), Neural Network based Genetic Algorithm (NNGA), and reinforcement learning and deep neural network (RLDNN). The IGPCA/CNN method outperforms other methods in energy prediction accuracy across short-term, medium-term, and long-term time scales. It achieves an MSE score of 0.564, lower than MBPNN, NNGA, and RLDNN. In medium-term prediction, it achieves an MSE score of 0.587, better than MBPNN, NNGA, and RLDNN. In long-term prediction, it achieves an MSE score of 0.629, lower than MBPNN and RLDNN. Further analysis and validation experiments are needed to ensure robustness and generalizability. Comparing the method with other state-of-the-art approaches can provide a comprehensive understanding of its superiority. The performance of the models is evaluated based on reliability and effectiveness in energy demand forecasting. The major conclusion of this study is that the proposed approach outperforms the initial models in accurately forecasting energy demand. The combination of DNNs and the improved Giza pyramid construction methodology results in enhanced performance, demonstrating superior reliability and effectiveness compared to other models. The study highlights the significance of accurate energy demand prediction for optimizing energy systems and reducing costs.

Enhancing Robustness of Medical Image Segmentation Model with Neural Memory Ordinary Differential Equation.

Deep neural networks (DNNs) have emerged as a prominent model in medical image segmentation, achieving remarkable advancements in clinical practice. Despite the promising results reported in the literature, the effectiveness of DNNs necessitates substantial quantities of high-quality annotated training data. During experiments, we observe a significant decline in the performance of DNNs on the test set when there exists disruption in the labels of the training dataset, revealing inherent limitations in the robustness of DNNs. In this paper, we find that the neural memory ordinary differential equation (nmODE), a recently proposed model based on ordinary differential equations (ODEs), not only addresses the robustness limitation but also enhances performance when trained by the clean training dataset. However, it is acknowledged that the ODE-based model tends to be less computationally efficient compared to the conventional discrete models due to the multiple function evaluations required by the ODE solver. Recognizing the efficiency limitation of the ODE-based model, we propose a novel approach called the nmODE-based knowledge distillation (nmODE-KD). The proposed method aims to transfer knowledge from the continuous nmODE to a discrete layer, simultaneously enhancing the model's robustness and efficiency. The core concept of nmODE-KD revolves around enforcing the discrete layer to mimic the continuous nmODE by minimizing the KL divergence between them. Experimental results on 18 organs-at-risk segmentation tasks demonstrate that nmODE-KD exhibits improved robustness compared to ODE-based models while also mitigating the efficiency limitation.

Divergences in color perception between deep neural networks and humans

Deep neural networks (DNNs) are increasingly proposed as models of human vision, bolstered by their impressive performance on image classification and object recognition tasks. Yet, the extent to which DNNs capture fundamental aspects of human vision such as color perception remains unclear. Here, we develop novel experiments for evaluating the perceptual coherence of color embeddings in DNNs, and we assess how well these algorithms predict human color similarity judgments collected via an online survey. We find that state-of-the-art DNN architectures — including convolutional neural networks and vision transformers — provide color similarity judgments that strikingly diverge from human color judgments of (i) images with controlled color properties, (ii) images generated from online searches, and (iii) real-world images from the canonical CIFAR-10 dataset. We compare DNN performance against an interpretable and cognitively plausible model of color perception based on wavelet decomposition, inspired by foundational theories in computational neuroscience. While one deep learning model — a convolutional DNN trained on a style transfer task — captures some aspects of human color perception, our wavelet algorithm provides more coherent color embeddings that better predict human color judgments compared to all DNNs we examine. These results hold when altering the high-level visual task used to train similar DNN architectures (e.g., image classification versus image segmentation), as well as when examining the color embeddings of different layers in a given DNN architecture. These findings break new ground in the effort to analyze the perceptual representations of machine learning algorithms and to improve their ability to serve as cognitively plausible models of human vision. Implications for machine learning, human perception, and embodied cognition are discussed.

Multi-exit DNN inference acceleration for intelligent terminal with heterogeneous processors

Recently, there has been a burgeoning popularity in the deployment of deep learning vision applications upon terminal devices. However, as the number of layers in deep neural networks (DNNs) and structural complexity increase, although the performance of DNN in handling computer vision tasks has become powerful, model inference tasks on computation resource constrained intelligent terminal devices are frequently incapable of meeting latency requirement. A commonly adopted solution to inference acceleration presents multi-exit DNNs to reduce latency via the provision of early exits. However, existing methods do not fully utilize the potential of heterogeneous processors (GPU/CPU) on intelligent terminal devices to cooperatively accelerate multi-exit DNN inference in parallel. Furthermore, the impact of complex image and video input on multi-exit DNNs, as well as the effects of different power consumption modes on processors within intelligent terminal devices, remain inadequately explored. To address these issues, we comprehensively considered the computing performance of heterogeneous processors in different power consumption modes, the structure and characteristic of multi-exit DNNs in inference acceleration, and proposed the Collaborative Inference Acceleration mechanism for intelligent terminal with Heterogeneous Processors (CIAHP). CIAHP includes a deep neural network computation time prediction model and a multi-exit DNN task allocation algorithm with heterogeneous processors. Our experiments demonstrate that CIAHP performs multi-exit DNN inference 2.31× faster than CPU alone, and is 1.23× faster than GPU alone when processing complex image samples.

A deep neural network approach combined with Findley parameter to predict fretting fatigue crack initiation lifetime

In the aviation field, the crack initiation stage could be the dominant one of the whole damage failure process of components subjecting to fretting fatigue. Thus, getting an accurate prediction of the crack initiation lifetime become an important topic in research of fretting fatigue problems. In this work, a new prediction method for fretting fatigue crack initiation lifetime is proposed, which is based on Deep Neural Network (DNN), Findley Parameter (FP) and Finite Element Method (FEM). Input features are shear stress amplitude and maximum normal stress in the critical plane, and crack initiation lifetime is set as the output target. Dropout and bootstrapping are considered to improve the performance of DNN. Through comparative analysis, it is shown that this new proposed approach can effectively predict the fretting fatigue crack initiation lifetime. The prediction accuracy and stability are greatly improved compared with the theoretical formulation based on FP.

EBDM: Ensemble binary detection models for multi-class wireless intrusion detection based on deep neural network

To effectively detect network intrusions, Machine Learning (ML) based detection models usually integrate Feature Selection (FS) methods for data enhancements as the ensemble models to achieve better detection accuracy (ACC). Deep Neural Network (DNN) based detection models can detect large of attacks better than ML-based models and DNN based models usually integrate with Stack Auto Encoder (SAE) as for data enhancements. Some models adopt Random Sampling (RS) as Data Balancing (DB) method to enhance the training data and improve ACC. Aegean Wi-Fi Intrusion Dataset (AWID) is a dedicated dataset for WiFi networks and labels three different attack types: impersonation, injection, and flooding. However, the multi-class detection models cannot detect different types of wireless attacks simultaneously with the customized detection configurations for each type of attacks. Thus, we propose Ensemble Binary Detection Model (EBDM) to convert the multi-class detection as binary detection models. Each Binary Model (BM) provides dedicated binary detection for each target class instead of several classes. EBDM collects the best detection results from BMs to achieve better ACC than the classical multi-class detection model. To improve ACC, EBDM proposes new data enhancement mechanisms for FS, DB, and SAE, with the individual detection configurations for the target class in each BM. We propose Ensemble FS (EFS) consisting of three feature scoring methods, Classification And Regression Tree, Random Forest, and Extra Tree, and two feature selection methods, Support Vector Machine and Logical Regression, to provide six selected feature sets and choose the best feature set based on ACC in each BM. EBDM proposes Ensemble DB consisting of Tomek Link, SMOTE, and RS to enhance typical DB adopting RS and provide customized sampling size by the target class size in each BM. EBDM provides four Encoder Modes for SAE to determine the best feature extraction configuration for each BM. EBDM integrates Squeeze-and-Excitation block with DNN to enhance DNN performance. Based on the improved BMs with the customized configurations, EBDM selects the best p-value from BMs. The experiment results show that EBDM can detect three different types of wireless attacks in AWID more accurately than the related works in AWID based on the classical multi-class detection model

Predicting ischemic stroke risk from atrial fibrillation based on multi-spectral fundus images using deep learning.

Ischemic stroke (IS) is one of the most common serious secondary diseases of atrial fibrillation (AF) within 1 year after its occurrence, both of which have manifestations of ischemia and hypoxia of the small vessels in the early phase of the condition. The fundus is a collection of capillaries, while the retina responds differently to light of different wavelengths. Predicting the risk of IS occurring secondary to AF, based on subtle differences in fundus images of different wavelengths, is yet to be explored. This study was conducted to predict the risk of IS occurring secondary to AF based on multi-spectrum fundus images using deep learning. A total of 150 AF participants without suffering from IS within 1 year after discharge and 100 IS participants with persistent arrhythmia symptoms or a history of AF diagnosis in the last year (defined as patients who would develop IS within 1 year after AF, based on fundus pathological manifestations generally prior to symptoms of the brain) were recruited. Fundus images at 548, 605, and 810 nm wavelengths were collected. Three classical deep neural network (DNN) models (Inception V3, ResNet50, SE50) were trained. Sociodemographic and selected routine clinical data were obtained. The accuracy of all DNNs with the single-spectral or multi-spectral combination images at the three wavelengths as input reached above 78%. The IS detection performance of DNNs with 605 nm spectral images as input was relatively more stable than with the other wavelengths. The multi-spectral combination models acquired a higher area under the curve (AUC) scores than the single-spectral models. The probability of IS secondary to AF could be predicted based on multi-spectrum fundus images using deep learning, and combinations of multi-spectrum images improved the performance of DNNs. Acquiring different spectral fundus images is advantageous for the early prevention of cardiovascular and cerebrovascular diseases. The method in this study is a beneficial preliminary and initiative exploration for diseases that are difficult to predict the onset time such as IS.

Boosting separated softmax with discrimination for class incremental learning

Deep neural networks (DNNs) suffer from catastrophic forgetting when learning new classes continually and tend to classify old samples to new classes. Existing methods like exemplar-memory and knowledge distillation alleviate forgetting, but face prediction bias due to data imbalance. Separated softmax (Ahn et al., 2021) was proposed to solve this problem. However, they ignore the discrimination between old classes and new classes, which greatly limits the performance of DNNs. To enhance model’s discrimination, we propose discriminative separated softmax. We divide new samples into two parts: one combined with old samples to build a balanced dataset of all seen classes, and the other combined with old samples as an imbalanced dataset. Furthermore, we apply mixup for these two datasets respectively to enhance discrimination. We evaluate our method on CIFAR100 and ImageNet100. Experimental results show that our method can effectively improve the discrimination and achieves superior performance compared with separated softmax, while outperforming state-of-the-art methods.

A two-stage data quality improvement strategy for deep neural networks in fault severity estimation

As fault diagnosis based on deep neural network (DNN) has been developing from laboratory to industrial application, creating the appropriate data pipeline to train and evaluate the models has become a noticeable challenge. For mechanical fault diagnosis, high-quality data should be consistent with comprehensive fault information. In this paper, a two-stage data quality improvement strategy is proposed to sculpt the data for deep learning models in fault severity estimation. In the first stage, the spatial information reconstruction (SIR) approach is developed to transform the sensing data from time domain into spatial domain to establish the relationship between fault symptom and spatial position. In the second stage, spatial domain signals are organized based on fault characteristics and types of DNN models, which aims to convert data into useful information. The proposed strategy has been verified by experimental cases, and a comprehensive evaluation on the performance of DNNs trained by different input data has been applied. The experimental results proved that with thoughtfully engineered data, baseline DNN models can achieve high accuracy and robustness under different speed conditions. This paper provides a way for engineering practitioners to design industrial DNN applications with their domain knowledge.

A deep neural network architecture for reliable 3D position and size determination for Lagrangian particle tracking using a single camera

Microfluidic flows feature typically fully three-dimensional velocity fields. However, often the optical access for measurements is limited. Astigmatism or defocus particle tracking velocimetry is a technique that enables the 3D position determination of individual particles by the analysis of astigmatic/defocused particle images. The classification and position determination of particles is a task well suited to deep neural networks (DNNs). In this work, two DNNs are used to extract the class and in-plane position (object detection) as well as the depth position (regression). The performance of both DNNs is assessed by the position uncertainties as well as the precision of the size classes and the amount of recalled particles. The DNNs are evaluated on a synthetic dataset and establish a new benchmark of DNNs in defocus tracking applications. The recall is higher than compared to classic methods and the in-plane errors are always subpixel accurate. The relative uncertainty in the depth position is below 1% for all examined particle seeding concentrations. Additionally, the performance on experimental images, using four different particle sizes, ranging from 1.14 to 5.03 is analyzed. The particle images are systematically rearranged to produce comprehensive datasets of varying particle seeding concentrations. The distinction between particles of similar size is more challenging but the DNNs still show very good results. A precision above 96% is reached with a high recall above 95%. The error in the depth position remains below 1% and the in-plane errors are subpixel accurate with respect to the labels. The work shows that first, DNNs can be trained with artificially rearranged data sets based on individual experimental images and are therefore easily adaptable to various experimental setups and applicable by non-experts. Second, the DNNs can be successfully adapted to determine additional variables as in this case the size of the suspended particles.

Optimizing DNNs Model Partitioning for Enhanced Performance on Edge Devices

Deep Neural Networks (DNNs) have proven effective in various applications due to their dominant performance. However, integrating DNNs into edge devices remains challenging due to the large size of the DNN model, which requires efficient model parallelization and workload partitioning. Previous attempts to address these challenges have focused on data and model parallelism but have fallen short in terms of finding the optimal DNN model partitions for efficient distribution, considering available resources. This paper presents a pipelined DNN model parallelism framework that improves the performance of DNNs on edge devices. The framework optimizes DNN model training by determining the optimal number of partitions based on available edge resources. This is achieved through a combination of data and model parallelism techniques, which efficiently distribute the workload across multiple processors to reduce training time. The framework also includes a task controller to manage computing resources effectively. The experimental results demonstrate the effectiveness of the proposed approach, showing a significant reduction in the model training time compared to a baseline model AlexNet.