Training Process Of Deep Neural Networks Research Articles

Towards Trustworthy Deep Learning

Deep neural networks (DNNs) have achieved unprecedented success across many scientific and engineering fields in the last decades. Despite its empirical success, unfortunately, recent studies have shown that there are various failure modes and blindspots in DNN models which may result in unexpected serious failures and potential harms, e.g. the existence of adversarial examples and small perturbations. This is not acceptable especially for safety critical and high stakes applications in the real-world, including healthcare, self-driving cars, aircraft control systems, hiring and malware detection protocols. Moreover, it has been challenging to understand why and when DNNs will fail due to their complicated structures and black-box behaviors. Lacking interpretability is one critical issue that may seriously hinder the deployment of DNNs in high-stake applications, which need interpretability to trust the prediction, to understand potential failures, and to be able to mitigate harms and eliminate biases in the model. To make DNNs trustworthy and reliable for deployment, it is necessary and urgent to develop methods and tools that can (i) quantify and improve their robustness against adversarial and natural perturbations, and (ii) understand their underlying behaviors and further correct errors to prevent injuries and damages. These are the important first steps to enable Trustworthy AI and Trustworthy Machine Learning. In this talk, I will survey a series of research efforts in my lab contributed to tackling the grand challenges in (i) and (ii). In the first part of my talk, I will overview our research effort in Robust Machine Learning since 2017, where we have proposed the first attack-agnostic robustness evaluation metric, the first efficient robustness certification algorithms for various types of perturbations, and efficient robust learning algorithms across supervised learning to deep reinforcement learning. In the second part of my talk, I will survey a series of exciting results in my lab on accelerating interpretable machine learning and explainable AI. Specifically, I will show how we could bring interpretability into deep learning by leveraging recent advances in multi-modal models. I'll present recent works in our group on automatically dissecting neural networks with open vocabulary concepts, designing interpretable neural networks without concept labels, and briefly overview our recent efforts on demystifying black-box DNN training process, automated neuron explanations for Large Language Models and the first robustness evaluation of a family of neuron-level interpretation techniques.

Rolling the dice for better deep learning performance: A study of randomness techniques in deep neural networks

This paper presents a comprehensive empirical investigation into the interactions between various randomization techniques in Deep Neural Networks (DNNs) and their impact on learning performance. It is well-established that injecting randomness into the training process of DNNs, through various approaches, at different stages, is often beneficial for reducing overfitting and improving generalization. Nonetheless, the interactions between randomness techniques such as weight noise, dropout, and many others remain poorly understood. Consequently, it is challenging to determine which methods can be effectively combined to optimize DNN performance. To address this issue, we categorize the existing randomness techniques into four key types: injection of noise/randomness at the data, model structure, optimization or learning stage. We use this classification to identify gaps in the current coverage of potential mechanisms for the introduction of randomness, leading to proposing two new techniques: adding noise to the loss function and random masking of the gradient updates.In our empirical study, we employ a Particle Swarm Optimizer (PSO) for hyperparameter optimization (HPO) to explore the space of possible configurations to determine where and how much randomness should be injected to maximize DNN performance. We assess the impact of various types and levels of randomness for DNN architectures across standard computer vision benchmarks: MNIST, FASHION-MNIST, CIFAR10, and CIFAR100. Across more than 30000 evaluated configurations, we perform a detailed examination of the interactions between randomness techniques and their combined impact on DNN performance. Our findings reveal that randomness through data augmentation and in weight initialization are the main contributors to performance improvement. Additionally, correlation analysis demonstrates that different optimizers, such as Adam and Gradient Descent with Momentum, prefer distinct types of randomization during the training process. A GitHub repository with the complete implementation and generated dataset is available.2

Get Your Foes Fooled: Proximal Gradient Split Learning for Defense Against Model Inversion Attacks on IoMT Data

The past decade has seen a rapid adoption of Artificial Intelligence (AI), specifically the deep learning networks, in Internet of Medical Things (IoMT) ecosystem. However, it has been shown recently that the deep learning networks can be exploited by adversarial attacks that not only make IoMT vulnerable to the data theft but also to the manipulation of medical diagnosis. The existing studies consider adding noise to the raw IoMT data or model parameters which not only reduces the overall performance concerning medical inferences but also is ineffective to the likes of deep leakage from gradients method. In this work, we propose proximal gradient split learning (PSGL) method for defense against the model inversion attacks. The proposed method intentionally attacks the IoMT data when undergoing the deep neural network training process at client side. We propose the use of proximal gradient method to recover gradient maps and a decision-level fusion strategy to improve the recognition performance. Extensive analysis show that the PGSL not only provides effective defense mechanism against the model inversion attacks but also helps in improving the recognition performance on publicly available datasets. We report 14.0$\%$, 17.9$\%$, and 36.9$\%$ gains in accuracy over reconstructed and adversarial attacked images, respectively.

Attack Classification of Imbalanced Intrusion Data for IoT Network Using Ensemble-Learning-Based Deep Neural Network

With the increase in popularity of Internet of Things (IoT) and rise in interconnected devices, the need to foster effective security mechanism to handle vulnerabilities and risks in IoT networks has become evident. Security mechanisms such as Intrusion Detection System (IDS) are designed and deployed in IoT network environment to ensure security and prevent unauthorized access to system and resources. Moreover, there have been efforts to design IDS using various Deep Learning (DL) techniques, as these techniques possess intriguing characteristic of representing data with high abstraction. However, intrusion detection datasets used in literature possess imbalance class distribution, which is one of the challenging issue in developing coherent and potent intrusion detection and classification system. In this paper, we aim to address class imbalance problem using ensemble learning approach, namely, Bagging classifier, that uses Deep Neural Network (DNN) as base estimator. Here, in the proposed approach, the training process of DNN is influenced by including class weights that advocates to create balanced training subsets for DNN. The desirability and merit of the proposed approach can be considered as two-fold as it aims to achieve generalization along with addressing the class imbalance problem in intrusion detection datasets. The performance of the proposed approach is evaluated using four intrusion detection datasets, namely, NSL-KDD, UNSW_NB-15, CIC-IDS-2017, and BoT-IoT. Result analysis of the proposed approach is illustrated using various evaluation metrics, namely, accuracy, precision, recall, f-score, and False Positive Rate (FPR). Moreover, results of the proposed approach are also statistically tested using Wilcoxon signed-rank test.

LHDNN: Maintaining High Precision and Low Latency Inference of Deep Neural Networks on Encrypted Data

The advancement of deep neural networks (DNNs) has prompted many cloud service providers to offer deep learning as a service (DLaaS) to users across various application domains. However, in current DLaaS prediction systems, users’ data are at risk of leakage. Homomorphic encryption allows operations to be performed on ciphertext without decryption, which can be applied to DLaaS to ensure users’ data privacy. However, mainstream homomorphic encryption schemes only support homomorphic addition and multiplication, and do not support the ReLU activation function commonly used in the activation layers of DNNs. Previous work used approximate polynomials to replace the ReLU activation function, but the DNNs they implemented either had low inference accuracy or high inference latency. In order to achieve low inference latency of DNNs on encrypted data while ensuring inference accuracy, we propose a low-degree Hermite deep neural network framework (called LHDNN), which uses a set of low-degree trainable Hermite polynomials (called LotHps) as activation layers of DNNs. Additionally, LHDNN integrates a novel weight initialization and regularization module into the LotHps activation layer, which makes the training process of DNNs more stable and gives a stronger generalization ability. Additionally, to further improve the model accuracy, we propose a variable-weighted difference training (VDT) strategy that uses ReLU-based models to guide the training of LotHps-based models. Extensive experiments on multiple benchmark datasets validate the superiority of LHDNN in terms of inference speed and accuracy on encrypted data.

EdgeMesh: A hybrid distributed training mechanism for heterogeneous edge devices

AbstractThe proliferation of large‐scale distributed Internet of Things (IoT) applications has resulted in a surge in demand for network models such as deep neural networks (DNNs) to be trained and inferred at the edge. Due to the central data transmission mechanism, heterogeneity of edge devices, and resource constraints, the existing single data‐parallel, and model‐parallel distributed training mechanisms frequently fail to fully utilize the computing power of edge devices, network topology and bandwidth resources. In light of the shortcomings mentioned earlier, this article proposes EdgeMesh, a hybrid parallel training mode based on the Mesh‐Tensorflow framework, consisting of an adaptive meshing strategy and a dynamic model convolutional partitioning strategy. The computing power of IoT edge devices significantly speeds up the DNN training process. In a resource‐constrained environment, each node only supports a subset of the model's parallel computing tasks, reducing communication and memory overhead while retaining high scalability. Experiments show that when compared to single‐machine training and data parallel mode, EdgeMesh distributed training mechanism can reduce the average delay by 3.2 times and average memory overhead by 43% while maintaining model accuracy. The computing power of IoT edge devices effectively accelerates the DNN training process.

A Non-Idealities Aware Software–Hardware Co-Design Framework for Edge-AI Deep Neural Network Implemented on Memristive Crossbar

In this work, a non-idealities aware software-hardware co-design framework for deep neural network (DNN) implemented on memristive crossbar is presented. The device level non-ideal factors such as device conductance variation, nonuniform quantization levels, device-to-device variation and programming failure probability are included in the model. At array level, the impact of line resistance and sneak path are considered using a new fast and accurate line resistance estimation model. The non-linearity and offset of the peripheral circuits are also considered. By incorporating these factors into a unified DNN training process, the neural network performance can be evaluated holistically. Furthermore, the proposed training process can effectively mitigate the impact of these non-idealities and reduce the inference accuracy degradations. Implemented in a hybrid fashion of Python and PyTorch, the proposed framework is evaluated with a simplified 5-layer VGG network implemented on measured <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$128\times 128$ </tex-math></inline-formula> RRAM array with 3-level weight resolution. For CIFAR-10 tasks, 83% inference accuracy is achieved with less than 3% accuracy drop compared to the ideal model.

A novel adaptive fault diagnosis algorithm for multi-machine equipment: application in bearing and diesel engine

This paper proposes an adaptive fault diagnosis algorithm based on vibration signals for fault diagnosis of bearings and diesel engines. First, the improved nonlinear gray wolf optimization algorithm (NGWO) is adopted to optimize the key parameter for variational mode decomposition (VMD) with the power spectral entropy as the fitness value. Meanwhile, adaptive noise reduction of the signal is realized. Then, sensitive fault features of bearings and diesel engines are selected through a feature sensitivity analysis on the vibration signals. Also, a single-layer sparse autoencoder is used to align the feature dimensions of each type of data to construct feature matrix samples. Subsequently, a deep neural network (DNN) consisting of a two-layer stacked sparse autoencoder (SSAE) and a Softmax classification layer is constructed to realize failure mode recognition. During the training process of DNN, a surrogate model formed by NGWO and a back propagation neural network is employed to optimize the hyperparameters of SSAE. Finally, to verify the effectiveness of the proposed fault diagnosis algorithm, fault diagnosis experiments are conducted on the fault data set of bearings and diesel engines. The diagnosis results show that the proposed method achieves high-precision fault diagnosis for bearings and diesel engines and performs stably for small samples.

BATUDE: Budget-Aware Neural Network Compression Based on Tucker Decomposition

Model compression is very important for the efficient deployment of deep neural network (DNN) models on resource-constrained devices. Among various model compression approaches, high-order tensor decomposition is particularly attractive and useful because the decomposed model is very small and fully structured. For this category of approaches, tensor ranks are the most important hyper-parameters that directly determine the architecture and task performance of the compressed DNN models. However, as an NP-hard problem, selecting optimal tensor ranks under the desired budget is very challenging and the state-of-the-art studies suffer from unsatisfied compression performance and timing-consuming search procedures. To systematically address this fundamental problem, in this paper we propose BATUDE, a Budget-Aware TUcker DEcomposition-based compression approach that can efficiently calculate optimal tensor ranks via one-shot training. By integrating the rank selecting procedure to the DNN training process with a specified compression budget, the tensor ranks of the DNN models are learned from the data and thereby bringing very significant improvement on both compression ratio and classification accuracy for the compressed models. The experimental results on ImageNet dataset show that our method enjoys 0.33% top-5 higher accuracy with 2.52X less computational cost as compared to the uncompressed ResNet-18 model. For ResNet-50, the proposed approach enables 0.37% and 0.55% top-5 accuracy increase with 2.97X and 2.04X computational cost reduction, respectively, over the uncompressed model.

Backdoor Attacks on Image Classification Models in Deep Neural Networks

Deep neural network (DNN) is applied widely in many applications and achieves state-of-the-art performance. However, DNN lacks transparency and interpretability for users in structure. Attackers can use this feature to embed trojan horses in the DNN structure, such as inserting a backdoor into the DNN, so that DNN can learn both the normal main task and additional malicious tasks at the same time. Besides, DNN relies on data set for training. Attackers can tamper with training data to interfere with DNN training process, such as attaching a trigger on input data. Because of defects in DNN structure and data, the backdoor attack can be a serious threat to the security of DNN. The DNN attacked by backdoor performs well on benign inputs while it outputs an attacker-specified label on trigger attached inputs. Backdoor attack can be conducted in almost every stage of the machine learning pipeline. Although there are a few researches in the backdoor attack on image classification, a systematic review is still rare in this field. This paper is a comprehensive review of backdoor attacks. According to whether attackers have access to the training data, we divide various backdoor attacks into two types: poisoning-based attacks and non-poisoning-based attacks. We go through the details of each work in the timeline, discussing its contribution and deficiencies. We propose a detailed mathematical backdoor model to summary all kinds of backdoor attacks. In the end, we provide some insights about future studies.

An Information Theoretic Interpretation to Deep Neural Networks.

With the unprecedented performance achieved by deep learning, it is commonly believed that deep neural networks (DNNs) attempt to extract informative features for learning tasks. To formalize this intuition, we apply the local information geometric analysis and establish an information-theoretic framework for feature selection, which demonstrates the information-theoretic optimality of DNN features. Moreover, we conduct a quantitative analysis to characterize the impact of network structure on the feature extraction process of DNNs. Our investigation naturally leads to a performance metric for evaluating the effectiveness of extracted features, called the H-score, which illustrates the connection between the practical training process of DNNs and the information-theoretic framework. Finally, we validate our theoretical results by experimental designs on synthesized data and the ImageNet dataset.

Visual vs internal attention mechanisms in deep neural networks for image classification and object detection

The so-called “attention mechanisms” in Deep Neural Networks (DNNs) denote an automatic adaptation of DNNs to capture representative features given a specific classification task and related data. Such attention mechanisms perform both globally by reinforcing feature channels and locally by stressing features in each feature map. Channel and feature importance are learnt in the global end-to-end DNNs training process. In this paper, we present a study and propose a method with a different approach, adding supplementary visual data next to training images. We use human visual attention maps obtained independently with psycho-visual experiments, both in task-driven or in free viewing conditions, or powerful models for prediction of visual attention maps. We add visual attention maps as new data alongside images, thus introducing human visual attention into the DNNs training and compare it with both global and local automatic attention mechanisms. Experimental results show that known attention mechanisms in DNNs work pretty much as human visual attention, but still the proposed approach allows a faster convergence and better performance in image classification tasks.

Theory of the Frequency Principle for General Deep Neural Networks

Along with fruitful applications of Deep Neural Networks (DNNs) to realistic problems, recently, some empirical studies of DNNs reported a universal phenomenon of Frequency Principle (F-Principle): a DNN tends to learn a target function from low to high frequencies during the training. The F-Principle has been very useful in providing both qualitative and quantitative understandings of DNNs. In this paper, we rigorously investigate the F-Principle for the training dynamics of a general DNN at three stages: initial stage, intermediate stage, and final stage. For each stage, a theorem is provided in terms of proper quantities characterizing the F-Principle. Our results are general in the sense that they work for multilayer networks with general activation functions, population densities of data, and a large class of loss functions. Our work lays a theoretical foundation of the F-Principle for a better understanding of the training process of DNNs.

Learning to optimize for resource allocation in LTE-U networks

This paper proposes a deep learning (DL) resource allocation framework to achieve the harmonious coexistence between the transceiver pairs (TPs) and the Wi-Fi users in LTE-U networks. The noncon- vex resource allocation is considered as a constrained learning problem and the deep neural network (DNN) is employed to approximate the optimal resource allocation decisions through unsupervised manner. A parallel DNN framework is proposed to deal with the two optimization variables in this problem, where one is the licensed power allocation unit and the other is the unlicensed time fraction occupied unit. Besides, to guarantee the feasibility of the proposed algorithm, the Lagrange dual method is used to relax the constraints into the DNN training process. Then, the dual variable and the DNN parameter are alternating update via the batch-based gradient decent method until the training process converges. Numerical results show that the proposed algorithm is feasible and has better performance than other general algorithms.

A Unified Framework for Cross-Domain and Cross-System Recommendations

Cross-Domain Recommendation (CDR) and Cross-System Recommendations (CSR) are two of the promising solutions to address the long-standing data sparsity problem in recommender systems. They leverage the relatively richer information, e.g., ratings, from the source domain or system to improve the recommendation accuracy in the target domain or system. Therefore, finding an accurate mapping of the latent factors across domains or systems is crucial to enhancing recommendation accuracy. However, this is a very challenging task because of the complex relationships between the latent factors of the source and target domains or systems. To this end, in this paper, we propose a Deep framework for both Cross-Domain and Cross-System Recommendations, called DCDCSR, based on Matrix Factorization (MF) models and a fully connected Deep Neural Network (DNN). Specifically, DCDCSR first employs the MF models to generate user and item latent factors and then employs the DNN to map the latent factors across domains or systems. More importantly, we take into account the rating sparsity degrees of individual users and items in different domains or systems and use them to guide the DNN training process for utilizing the rating data more effectively. Extensive experiments conducted on three real-world datasets demonstrate that DCDCSR framework outperforms the state-of-the-art CDR and CSR approaches in terms of recommendation accuracy.

A rolling bearing fault diagnosis method based on fastDTW and an AGBDBN

In order to improve fault feature extraction and diagnosis for rolling bearings, a fault diagnosis method based on fast dynamic time warping (fastDTW) and an adaptive Gaussian-Bernoulli deep belief network (AGBDBN) is proposed in this paper. Firstly, for the non-stationary vibration signal characteristics of the bearing, the fastDTW algorithm is used to calculate the residual vector of the fault signal, thereby enhancing the fault characteristic information. Then, according to the continuous vibration value of the bearing vibration signal, a standard deep belief network (DBN) is improved to deal with the problem that the optimal setting for the learning rate is difficult to achieve in the deep neural network training process and the AGBDBN model is used for fault diagnosis. Finally, the proposed method is compared with a variety of model diagnosis methods. The experimental results show that the proposed method achieved good diagnostic results.

A survey of swarm and evolutionary computing approaches for deep learning

Deep learning (DL) has become an important machine learning approach that has been widely successful in many applications. Currently, DL is one of the best methods of extracting knowledge from large sets of raw data in a (nearly) self-organized manner. The technical design of DL depends on the feed-forward information flow principle of artificial neural networks with multiple layers of hidden neurons, which form deep neural networks (DNNs). DNNs have various architectures and parameters and are often developed for specific applications. However, the training process of DNNs can be prolonged based on the application and training set size (Gong et al. 2015). Moreover, finding the most accurate and efficient architecture of a deep learning system in a reasonable time is a potential difficulty associated with this approach. Swarm intelligence (SI) and evolutionary computing (EC) techniques represent simulation-driven non-convex optimization frameworks with few assumptions based on objective functions. These methods are flexible and have been proven effective in many applications; therefore, they can be used to improve DL by optimizing the applied learning models. This paper presents a comprehensive survey of the most recent approaches involving the hybridization of SI and EC algorithms for DL, the architecture of DNNs, and DNN training to improve the classification accuracy. The paper reviews the significant roles of SI and EC in optimizing the hyper-parameters and architectures of a DL system in context to large scale data analytics. Finally, we identify some open problems for further research, as well as potential issues related to DL that require improvements, and an extensive bibliography of the pertinent research is presented.

Accelerating Deep Learning with a Parallel Mechanism Using CPU + MIC

Deep neural networks (DNNs) is one of the most popular machine learning methods and is widely used in many modern applications. The training process of DNNs is a time-consuming process. Accelerating the training of DNNs has been the focus of many research works. In this paper, we speed up the training of DNNs applied for automatic speech recognition and the target architecture is heterogeneous (CPU + MIC). We apply asynchronous methods for I/O and communication operations and propose an adaptive load balancing method. Besides, we also employ a momentum idea to speed up the convergence of the gradient descent algorithm. Experimental results show that our optimized algorithm is able to acquire a 20-fold speedup on a CPU + MIC platform compared with the original sequential algorithm on a single-core CPU.

Wind Turbine Gearbox Failure Identification With Deep Neural Networks

The feasibility of monitoring the health of wind turbine (WT) gearboxes based on the lubricant pressure data in the supervisory control and data acquisition system is investigated in this paper. A deep neural network (DNN)-based framework is developed to monitor conditions of WT gearboxes and identify their impending failures. Six data-mining algorithms, the k- nearest neighbors, least absolute shrinkage and selection operator, ridge regression (Ridge), support vector machines, shallow neural network, as well as DNN, are applied to model the lubricant pressure. A comparative analysis of developed data-driven models is conducted and the DNN model is the most accurate. To prevent the overfitting of the DNN model, a dropout algorithm is applied into the DNN training process. Computational results show that the prediction error will shift before the occurrences of gearbox failures. An exponentially weighted moving average control chart is deployed to derive criteria for detecting the shifts. The effectiveness of the proposed monitoring approach is demonstrated by examining real cases from wind farms in China and benchmarked against the gearbox monitoring based on the oil temperature data.

Speech dereverberation for enhancement and recognition using dynamic features constrained deep neural networks and feature adaptation

This paper investigates deep neural networks (DNN) based on nonlinear feature mapping and statistical linear feature adaptation approaches for reducing reverberation in speech signals. In the nonlinear feature mapping approach, DNN is trained from parallel clean/distorted speech corpus to map reverberant and noisy speech coefficients (such as log magnitude spectrum) to the underlying clean speech coefficients. The constraint imposed by dynamic features (i.e., the time derivatives of the speech coefficients) are used to enhance the smoothness of predicted coefficient trajectories in two ways. One is to obtain the enhanced speech coefficients with a least square estimation from the coefficients and dynamic features predicted by DNN. The other is to incorporate the constraint of dynamic features directly into the DNN training process using a sequential cost function. In the linear feature adaptation approach, a sparse linear transform, called cross transform, is used to transform multiple frames of speech coefficients to a new feature space. The transform is estimated to maximize the likelihood of the transformed coefficients given a model of clean speech coefficients. Unlike the DNN approach, no parallel corpus is used and no assumption on distortion types is made. The two approaches are evaluated on the REVERB Challenge 2014 tasks. Both speech enhancement and automatic speech recognition (ASR) results show that the DNN-based mappings significantly reduce the reverberation in speech and improve both speech quality and ASR performance. For the speech enhancement task, the proposed dynamic feature constraint help to improve cepstral distance, frequency-weighted segmental signal-to-noise ratio (SNR), and log likelihood ratio metrics while moderately degrades the speech-to-reverberation modulation energy ratio. In addition, the cross transform feature adaptation improves the ASR performance significantly for clean-condition trained acoustic models.