Training Deep Neural Networks Research Articles

Early stage black pepper leaf disease prediction based on transfer learning using ConvNets

Plants get exposed to diseases, insects and fungus. This causes heavy damages to crop resulting in various leaves diseases. Leaf diseases can be diagnosed at an early stage with the aid of a smart computer vision system and timely disease prevention can be targeted. Black pepper is a medicinal plant that is extensively used in Ayurvedic medicine because of its therapeutic properties. The proposed work represents an intelligent transfer learning technique through state-of-the-art deep learning implementation using convolutional neural network to predict the presence of prominent diseases in black pepper leaves. The ImageNet dataset available online is used for training deep neural network. Later, this trained network is utilized for the prediction of the newly developed black pepper leaf image dataset. The developed data set consist of real time leaf images, which are candidly taken from the fields and annotated under supervision of an expert. The leaf diseases considered are anthracnose, slow wilt, early stage phytophthora, phytophthora and yellowing. The hyperparameters chosen for tuning in to deep learning models are initial learning rates, optimization algorithm, image batches, epochs, validation and training data, etc. The accuracy obtained with 0.001 learning rate ranges from 99.1 to 99.7% for the Inception V3, GoogleNet, SqueezeNet and Resnet18 models. Proposed Resnet18 model outperforms all model with 99.67% accuracy. The resulting validation accuracy obtained using these models is high and the validation loss is low. This work represents improvement in agriculture and a cutting edge deep neural network method for early stage leaf disease identification and prediction. This is an approach using a deep learning network to predict early stage black pepper leaf diseases.

Ultra-short-term wind power probabilistic forecasting based on an evolutionary non-crossing multi-output quantile regression deep neural network

Ultra-short-term wind power probabilistic forecasting is of significance for stable power grid operation; however, it is still challenging due to the inherent nonlinearity and uncertainty. Most state-of-the-art methods have focused on achieving quantile prediction using a combination of linear quantile regression and nonlinear complex deep neural networks. Yet, these methods struggle with several dilemmas. Quantile regression deep neural networks require a complete training once for each quantile. The multi-training mode and complex structure of quantile regression deep neural network can lead to extremely high computational complexity. Most of the training of quantile regression deep neural networks are guided by the loss of each quantile, and the weights are adjusted by gradient descent in which the gradient explosion and quantile crossover may be encountered. Therefore, this paper proposes a non-crossing multi-output quantile regression deep neural network optimized by chaotic particle swarm optimization. It designs a multi-output deep neural network to output all quantile estimations simultaneously through one training, effectively solving the structural complexity problem of traditional quantile regression deep neural networks. Since quantile regression produces a non-differentiable loss function which significantly hinders model training, the proposed neural network is trained by chaotic particle swarm optimization. It not only achieves the effect of optimizing all quantile losses simultaneously, but also can significantly alleviate the dilemma of training in traditional neural network weight optimization. In addition, several non-crossing constraints are designed for avoiding quantile crossover. The proposed model is trained and tested on two real-world wind power case studies. The experiment results show that the proposed model shows superiority in performance criteria, training speed, and avoiding quantile crossover.

Multi-modal brain tumor segmentation via conditional synthesis with Fourier domain adaptation

Accurate brain tumor segmentation is critical for diagnosis and treatment planning, whereby multi-modal magnetic resonance imaging (MRI) is typically used for analysis. However, obtaining all required sequences and expertly labeled data for training is challenging and can result in decreased quality of segmentation models developed through automated algorithms.In this work, we examine the possibility of employing a conditional generative adversarial network (GAN) approach for synthesizing multi-modal images to train deep learning-based neural networks aimed at high-grade glioma (HGG) segmentation. The proposed GAN is conditioned on auxiliary brain tissue and tumor segmentation masks, allowing us to attain better accuracy and control of tissue appearance during synthesis. To reduce the domain shift between synthetic and real MR images, we additionally adapt the low-frequency Fourier space components of synthetic data, reflecting the style of the image, to those of real data. We demonstrate the impact of Fourier domain adaptation (FDA) on the training of 3D segmentation networks and attain significant improvements in both the segmentation performance and prediction confidence. Similar outcomes are seen when such data is used as a training augmentation alongside the available real images. In fact, experiments on the BraTS2020 dataset reveal that models trained solely with synthetic data exhibit an improvement of up to 4% in Dice score when using FDA, while training with both real and FDA-processed synthetic data through augmentation results in an improvement of up to 5% in Dice compared to using real data alone. This study highlights the importance of considering image frequency in generative approaches for medical image synthesis and offers a promising approach to address data scarcity in medical imaging segmentation.

Analog Resistive Switching Devices for Training Deep Neural Networks with the Novel Tiki-Taka Algorithm.

A critical bottleneck for the training of large neural networks (NNs) is communication with off-chip memory. A promising mitigation effort consists of integrating crossbar arrays of analogue memories in the Back-End-Of-Line, to store the NN parameters and efficiently perform the required synaptic operations. The "Tiki-Taka" algorithm was developed to facilitate NN training in the presence of device nonidealities. However, so far, a resistive switching device exhibiting all the fundamental Tiki-Taka requirements, which are many programmable states, a centered symmetry point, and low programming noise, was not yet demonstrated. Here, a complementary metal-oxide semiconductor (CMOS)-compatible resistive random access memory (RRAM), showing more than 30 programmable states with low noise and a symmetry point with only 5% skew from the center, is presented for the first time. These results enable generalization of Tiki-Taka training from small fully connected networks to larger long-/short-term-memory types of NN.

Better schedules for low precision training of deep neural networks

Better schedules for low precision training of deep neural networks

On the ideal number of groups for isometric gradient propagation

Recently, various normalization layers have been proposed to stabilize the training of deep neural networks. Among them, group normalization is a generalization of layer normalization and instance normalization by allowing a degree of freedom in the number of groups it uses. However, to determine the optimal number of groups, trial-and-error-based hyperparameter tuning is required, and such experiments are time-consuming. In this study, we discuss a reasonable method for setting the number of groups. First, we find that the number of groups influences the gradient behavior of the group normalization layer. Based on this observation, we derive the ideal number of groups, which calibrates the gradient scale to facilitate gradient descent optimization. This paper is the first to propose an optimal number of groups that is theoretically grounded, architecture-aware, and can provide a proper value in a layer-wise manner for all layers. The proposed method exhibited improved performance over existing methods in numerous neural network architectures, tasks, and datasets.

Calibration-free quantitative phase imaging in multi-core fiber endoscopes using end-to-end deep learning.

Quantitative phase imaging (QPI) through multi-core fibers (MCFs) has been an emerging in vivo label-free endoscopic imaging modality with minimal invasiveness. However, the computational demands of conventional iterative phase retrieval algorithms have limited their real-time imaging potential. We demonstrate a learning-based MCF phase imaging method that significantly reduced the phase reconstruction time to 5.5 ms, enabling video-rate imaging at 181 fps. Moreover, we introduce an innovative optical system that automatically generated the first, to the best of our knowledge, open-source dataset tailored for MCF phase imaging, comprising 50,176 paired speckles and phase images. Our trained deep neural network (DNN) demonstrates a robust phase reconstruction performance in experiments with a mean fidelity of up to 99.8%. Such an efficient fiber phase imaging approach can broaden the applications of QPI in hard-to-reach areas.

Image Signal Processing in the Context of Deep Learning Applications

Deep learning accelerators are a specialized sort of hardware architecture designed to enhance the computational efficiency of computers engaged in deep neural networks (DNNs) training. The implementation of DNNs in embedded vision applications might potentially be facilitated by the integration of energy-effective accelerators of deep learning into sensors. The lack of recognition for their significant impact on accuracy is a notable oversight. In previous iterations of deep learning accelerators integrated inside sensors, a common approach was bypassing the image signal processor (ISP). This deviation from the traditional vision pipelines had a detrimental impact on the performance of machine learning models trained on data that had undergone post-ISP processing. In this study, we establish a set of energy-efficient techniques that allow ISP to maximize their advantages while also limiting the covariate shift between the target dataset (RAW images) and the training dataset (ISP-analyzed images). This approach enables the practical use of in-sensor accelerators. To clarify, our results do not minimize the relevance of in-sensor accelerators. Instead, we highlight deficiencies in the methodology used in prior research and propose methodologies that empower in-sensor accelerators to fully exploit their capabilities.

Missing Data Imputation Using Socio-hawk Optimization-based Deep Neural Network

Missing data remain the common issue experienced in the real-world environment, which leads to deviation in data analysis and mining. Therefore, in order to lessen the consequences of missing data caused by human mistake, missing data imputation must be used in data processing. The traditional imputation model fails to satisfy the evaluation requirement due to its poor stability and low accuracy. Further, these models compromise the imputation accuracy of the increasing number of missing information. Hence, in this research, an optimized missing data imputation model is proposed using the Socio-hawk optimization Deep Neural Network (DNN). In this research, the DNN extracts the important features from the data, in which the missing data are estimated with an arbitrary missing pattern. It is stated that whenever the hyperparameters are tuned properly, the DNN’s performance is improved. The key here is the efficient training of DNN using the suggested Socio-hawk optimization, which improves the imputation model’s accuracy. To determine how well the suggested imputation model imputes missing data, it is compared to other methods. As a result, the paper’s primary contribution is to effectively train DNN using the suggested Socio-hawk optimization that reduces the error rate of the imputation model. The experimental evaluation shows that the proposed missing data imputation model attains a high performance at 90%, which provides 1.0595, 1.9919, and 0.9421 of MAE, MSE, and MAPE.

Research on 3D Geological Modeling Method Based on Deep Neural Networks for Drilling Data

Three-dimensional (3D) models provide the most intuitive representation of geological conditions. Traditional modeling methods heavily depend on technicians’ expertise and lack ease of updating. In this study, we introduce a deep learning-based method for 3D geological implicit modeling, leveraging a substantial dataset of geological drilling data. By applying resampling and normalization techniques, we standardize drilling data and significantly expand the dataset, making it suitable for training deep neural networks. Utilizing the characteristics of the sample data, we design and establish the network structure, loss function, and parameter configurations, resulting in the training of a deep neural network with high accuracy and robust generalization capability. Ultimately, we utilize the dataset generated from the network’s predictions to render and construct the 3D geological model. The research in this paper demonstrates the significant promise of deep neural networks in addressing geological challenges. The deep learning-based implicit 3D modeling method surpasses traditional approaches in terms of generalization, convenience, and adaptability.

Instance segmentation on distributed deep learning big data cluster

AbstractDistributed deep learning is a promising approach for training and deploying large and complex deep learning models. This paper presents a comprehensive workflow for deploying and optimizing the YOLACT instance segmentation model as on big data clusters. OpenVINO, a toolkit known for its high-speed data processing and ability to optimize deep learning models for deployment on a variety of devices, was used to optimize the YOLACT model. The model is then run on a big data cluster using BigDL, a distributed deep learning library for Apache Spark. BigDL provides a high-level programming interface for defining and training deep neural networks, making it suitable for large-scale deep learning applications. In distributed deep learning, input data is divided and distributed across multiple machines for parallel processing. This approach offers several advantages, including the ability to handle very large data that can be stored in a distributed manner, scalability to decrease processing time by increasing the number of workers, and fault tolerance. The proposed workflow was evaluated on virtual machines and Azure Databricks, a cloud-based platform for big data analytics. The results indicated that the workflow can scale to large datasets and deliver high performance on Azure Databricks. This study explores the benefits and challenges of using distributed deep learning on big data clusters for instance segmentation. Popular distributed deep learning frameworks are discussed, and BigDL is chosen. Overall, this study highlights the practicality of distributed deep learning for deploying and scaling sophisticated deep learning models on big data clusters.

A Generic, High-Performance, Compression-Aware Framework for Data Parallel DNN Training

Gradient compression is a promising approach to alleviating the communication bottleneck in data parallel deep neural network (DNN) training by significantly reducing the data volume of gradients for synchronization. While gradient compression is being actively adopted by the industry (e.g., Facebook and AWS), our study reveals that there are two critical but often overlooked challenges: 1) inefficient coordination between compression and communication during gradient synchronization incurs substantial overheads, and 2) developing, optimizing, and integrating gradient compression algorithms into DNN systems imposes heavy burdens on DNN practitioners, and ad-hoc compression implementations often yield surprisingly poor system performance. In this paper, we propose a compression-aware gradient synchronization architecture, <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CaSync</monospace> , which relies on flexible composition of basic computing and communication primitives. It is general and compatible with any gradient compression algorithms and gradient synchronization strategies and enables high-performance computation-communication pipelining. We further introduce a gradient compression toolkit, <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CompLL</monospace> , to enable efficient development and automated integration of on-GPU compression algorithms into DNN systems with little programming burden. Lastly, we build a compression-aware DNN training framework <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HiPress</monospace> with <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CaSync</monospace> and <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CompLL</monospace> . <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HiPress</monospace> is open-sourced and runs on mainstream DNN systems such as MXNet, TensorFlow, and PyTorch. Evaluation via a 16-node cluster with 128 NVIDIA V100 GPUs and a 100 Gbps network shows that <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HiPress</monospace> improves the training speed over current compression-enabled systems (e.g., BytePS-onebit, Ring-DGC and PyTorch-PowerSGD) by 9.8%-69.5% across six popular DNN models.

An In-Situ Visual Analytics Framework for Deep Neural Networks.

The past decade has witnessed the superior power of deep neural networks (DNNs) in applications across various domains. However, training a high-quality DNN remains a non-trivial task due to its massive number of parameters. Visualization has shown great potential in addressing this situation, as evidenced by numerous recent visualization works that aid in DNN training and interpretation. These works commonly employ a strategy of logging training-related data and conducting post-hoc analysis. Based on the results of offline analysis, the model can be further trained or fine-tuned. This strategy, however, does not cope with the increasing complexity of DNNs, because (1) the time-series data collected over the training are usually too large to be stored entirely; (2) the huge I/O overhead significantly impacts the training efficiency; (3) post-hoc analysis does not allow rapid human-interventions (e.g., stop training with improper hyper-parameter settings to save computational resources). To address these challenges, we propose an in-situ visualization and analysis framework for the training of DNNs. Specifically, we employ feature extraction algorithms to reduce the size of training-related data in-situ and use the reduced data for real-time visual analytics. The states of model training are disclosed to model designers in real-time, enabling human interventions on demand to steer the training. Through concrete case studies, we demonstrate how our in-situ framework helps deep learning experts optimize DNNs and improve their analysis efficiency.

CROMOSim: A Deep Learning-Based Cross-Modality Inertial Measurement Simulator

With the prevalence of wearable devices, inertial measurement unit (IMU) data has been utilized in monitoring and assessing human mobility such as human activity recognition (HAR) and human pose estimation (HPE). Training deep neural network (DNN) models for these tasks require a large amount of labelled data, which are hard to acquire in uncontrolled environments. To mitigate the data scarcity problem, we design CROMOSim, a cross-modality sensor simulator that simulates high fidelity virtual IMU sensor data from motion capture systems or monocular RGB cameras. It utilizes a skinned multi-person linear model (SMPL) for 3D body pose and shape representations to enable simulation from arbitrary on-body positions. Then a DNN model is trained to learn the functional mapping from imperfect trajectory estimations in a 3D SMPL body tri-mesh due to measurement noise, calibration errors, occlusion and other modelling artifacts, to IMU data. We evaluate the fidelity of CROMOSim simulated data and its utility in data augmentation on various HAR and HPE datasets. Extensive empirical results show that the proposed model achieves a 6.7% improvement over baseline methods in a HAR task.

Deep Learning Based Detection of Toxic Mushrooms in Karnataka

Mushrooms are fungi that play significant roles in ecosystems, food, and medicine. However, distinguishing toxic mushrooms from non-toxic ones is challenging, as their visual characteristics can be misleading. Deep learning, a subset of artificial intelligence, offers a promising approach to detecting and classifying mushrooms based on their toxicity. By training deep neural networks on datasets of labelled mushroom images, it becomes possible to create models capable of accurately differentiating between toxic and non-toxic mushrooms. Mushrooms are fungi that play significant roles in ecosystems, food, and medicine. However, manually distinguishing toxic mushrooms from non-toxic ones is a challenging and time-consuming task. This research work is to help the local farmers and exporters to identify the difference between the toxic and the edible mushroom which will eventually boost production and will directly have an impact on the economic growth in India. This paper compared different deep learning algorithms such as Support Vector Machine (SVM), Resnet50, YOLO V5 and Alexnet, which have been found to be the most useful in classifying the mushrooms effectively as toxic and non-toxic respectively. The patterns in the images of the mushrooms have been predicted through the above classifiers. The validation techniques show SVM to be the most effective with an accuracy of 83%.

Stable tensor neural networks for efficient deep learning.

Learning from complex, multidimensional data has become central to computational mathematics, and among the most successful high-dimensional function approximators are deep neural networks (DNNs). Training DNNs is posed as an optimization problem to learn network weights or parameters that well-approximate a mapping from input to target data. Multiway data or tensors arise naturally in myriad ways in deep learning, in particular as input data and as high-dimensional weights and features extracted by the network, with the latter often being a bottleneck in terms of speed and memory. In this work, we leverage tensor representations and processing to efficiently parameterize DNNs when learning from high-dimensional data. We propose tensor neural networks (t-NNs), a natural extension of traditional fully-connected networks, that can be trained efficiently in a reduced, yet more powerful parameter space. Our t-NNs are built upon matrix-mimetic tensor-tensor products, which retain algebraic properties of matrix multiplication while capturing high-dimensional correlations. Mimeticity enables t-NNs to inherit desirable properties of modern DNN architectures. We exemplify this by extending recent work on stable neural networks, which interpret DNNs as discretizations of differential equations, to our multidimensional framework. We provide empirical evidence of the parametric advantages of t-NNs on dimensionality reduction using autoencoders and classification using fully-connected and stable variants on benchmark imaging datasets MNIST and CIFAR-10.

Intermittent Deployment for Large-Scale Multi-Robot Forage Perception: Data Synthesis, Prediction, and Planning

Monitoring the health and vigor of grasslands is vital for informing management decisions to optimize rotational grazing in agriculture applications. To take advantage of forage resources and improve land productivity, we require knowledge of pastureland growth patterns that is simply unavailable at the state of the art. In this paper, we propose to deploy a team of robots to monitor the evolution of an unknown pastureland environment to fulfill the above goal. To monitor such an environment, which usually evolves slowly, we need to design a strategy for rapid assessment of the environment over large areas at a low cost. Thus, we propose an integrated pipeline comprising data synthesis, deep neural network training, and prediction along with a multi-robot deployment algorithm that monitors pasturelands intermittently. Specifically, using expert-informed agricultural data coupled with novel data synthesis in ROS Gazebo, we first propose a new neural network architecture to learn the spatiotemporal dynamics of the environment. Such predictions help us to understand pastureland growth patterns on large scales and make appropriate monitoring decisions for the future. Based on our predictions, we then design an intermittent multi-robot deployment policy for low-cost monitoring. Finally, we compare the proposed pipeline with other methods, from data synthesis to prediction and planning, to corroborate our pipeline’s performance. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —Pasturelands are an integral part of agricultural production in the United States. To take full advantage of the forage resource and avoid environmental degradation, pastureland must be managed optimally. This paper focuses on the question of how to deploy robot teams to sense and model physical processes over varying timescales. The goal of this work is to develop a new integrated pipeline for the long-term deployment of heterogeneous robot teams grounded in the problem of autonomous monitoring in precision grazing to improve land productivity. By using the proposed pipeline in grassland ecosystem management, we will have a better understanding of the physical environment while respecting energy budgets.

Sign-Based Gradient Descent With Heterogeneous Data: Convergence and Byzantine Resilience.

Communication overhead has become one of the major bottlenecks in the distributed training of modern deep neural networks. With such consideration, various quantization-based stochastic gradient descent (SGD) solvers have been proposed and widely adopted, among which signSGD with majority vote shows a promising direction because of its communication efficiency and robustness against Byzantine attackers. However, signSGD fails to converge in the presence of data heterogeneity, which is commonly observed in the emerging federated learning (FL) paradigm. In this article, a sufficient condition for the convergence of the sign-based gradient descent method is derived, based on which a novel magnitude-driven stochastic-sign-based gradient compressor is proposed to address the non-convergence issue of signSGD. The convergence of the proposed method is established in the presence of arbitrary data heterogeneity. The Byzantine resilience of sign-based gradient descent methods is quantified, and the error-feedback mechanism is further incorporated to boost the learning performance Experimental results on the MNIST dataset, the CIFAR-10 dataset, and the Tiny-ImageNet dataset corroborate the effectiveness of the proposed methods.

AFEX-Net: Adaptive feature extraction convolutional neural network for classifying computerized tomography images.

Deep convolutional neural networks are favored methods that are widely used in medical image processing due to their demonstrated performance in this area. Recently, the emergence of new lung diseases, such as COVID-19, and the possibility of early detection of their symptoms from chest computerized tomography images has attracted many researchers to classify diseases by training deep convolutional neural networks on lung computerized tomography images. The trained networks are expected to distinguish between different lung indications in various diseases, especially at the early stages. The purpose of this study is to introduce and assess an efficient deep convolutional neural network, called AFEX-Net, that can classify different lung diseases from chest computerized tomography images. We designed a lightweight convolutional neural network called AFEX-Net with adaptive feature extraction layers, adaptive pooling layers, and adaptive activation functions. We trained and tested AFEX-Net on a dataset of more than 10,000 chest computerized tomography slices from different lung diseases (CC dataset), using an effective pre-processing method to remove bias. We also applied AFEX-Net to the public COVID-CTset dataset to assess its generalizability. The study was mainly conducted based on data collected over approximately six months during the pandemic outbreak in Afzalipour Hospital, Iran, which is the largest hospital in Southeast Iran. AFEX-Net achieved high accuracy and fast training on both datasets, outperforming several state-of-the-art convolutional neural networks. It has an accuracy of and on the CC and COVID-CTset datasets, respectively, with a learning speed that is 3 times faster compared to similar methods due to its lightweight structure. AFEX-Net was able to extract distinguishing features and classify chest computerized tomography images, especially at the early stages of lung diseases. The AFEX-Net is a high-performing convolutional neural network for classifying lung diseases from chest CT images. It is efficient, adaptable, and compatible with input data, making it a reliable tool for early detection and diagnosis of lung diseases.

Deep learning framework for bovine iris segmentation.

Iris segmentation is an initial step for identifying the biometrics of animals when establishing a traceability system for livestock. In this study, we propose a deep learning framework for pixel-wise segmentation of bovine iris with a minimized use of annotation labels utilizing the BovineAAEyes80 public dataset. The proposed image segmentation framework encompasses data collection, data preparation, data augmentation selection, training of 15 deep neural network (DNN) models with varying encoder backbones and segmentation decoder DNNs, and evaluation of the models using multiple metrics and graphical segmentation results. This framework aims to provide comprehensive and in-depth information on each model's training and testing outcomes to optimize bovine iris segmentation performance. In the experiment, U-Net with a VGG16 backbone was identified as the optimal combination of encoder and decoder models for the dataset, achieving an accuracy and dice coefficient score of 99.50% and 98.35%, respectively. Notably, the selected model accurately segmented even corrupted images without proper annotation data. This study contributes to the advancement of iris segmentation and the establishment of a reliable DNN training framework.