Search

This paper addresses the questions and concerns raised in the comment paper entitled “Comments on ‘Traffic Sign Recognition Using Kernel Extreme Leaning Machines With Deep Perceptual Features”’. It is analyzed that the major questions and concerns in the comment paper are due to some misunderstanding on the statements and the main contributions of the original paper by Zeng et al.. The main clarifications are as follows: (1) The weight update rule used by Zeng et al. is a standard form of kernel extreme learning machine (KELM) but not the rule of reduced KELM mentioned by the comment paper. (2) The aim and main contribution of the original paper by Zeng et al. are to improve the classification precision of traffic sign recognition (TSR) based on convolutional neural networks without dramatically increasing the scale or complexity of the deep neural network model, which leads to a relatively less training cost. The authors of the comment paper focused on the reduction of KELM’s computational costs but it is beyond the topic of the original paper. In addition, the reduction of KELM’s computational costs has been well studied in the literature. (3) The supplementary experiment is provided to show that the TSR method proposed by Zeng et al. is valid whether or not the training samples are much more than the needed kernels. [ABSTRACT FROM AUTHOR]

Surface defect is a serious problem in real-world manufacturing system and it is important to use vision-based recognition to ensure the surface quality of products. Currently, due to the ability of automatic feature extraction, deep learning models, such as convolutional neural network (CNN), have been widely used in this area. However, these CNN-based models may not solve a problem well - inter-class similarities and intra-class variations (ISIV), which affect their ability of feature extraction and thus influence their recognition performance. To address this problem, this paper introduces a graph guidance mechanism into CNN to improve the ability of feature extraction, called Graph guided Convolutional Neural Network (GCNN). Firstly, GCNN defines a graph by computing the similarities between training samples. Secondly, the graph is introduced into VGG11, a popular CNN structure, to increase the inter-class distances and decrease the intra-class distances between defect samples. Meanwhile, a learnable coefficient is introduced into the training process to balance the effect of graph guidance automatically. The experimental results on four famous surface defect datasets demonstrate that the graph guidance helps CNN models have better ability of feature extraction and thus achieve better performance. Compared with state-of-the-art models, the proposed method can achieve the best performance. Furthermore, the final discussion shows that the learnable coefficient can help the proposed model to gain better performance, and that the proposed model increases a little computation cost compared to its original CNN model. Note to Practitioners—This paper is motivated by the problem in real-world manufacturing process – inter-class similarities and intra-class variations. Most of current CNN-based models may not solve this problem well, which limits their applications in real-world system. This paper proposes a graph guidance mechanism and introduces it to the training of CNN. The proposal can improve the ability of feature extraction, and thus have better performance than those without the graph guidance. By applying the proposal to four examples, the results show that the proposal is more feasible and effective than the state-of-the-art models in surface defect recognition. Furthermore, the mechanism assists the training of CNN and is removed in the process of recognition, so it will not increase time and occupied memory in recognition, and meanwhile, only limited computation cost is increased in training process. [ABSTRACT FROM AUTHOR]

This paper investigates the feasibility of novel data-driven deep learning (DL) models to predict the solution of Maxwell’s equations for low-frequency electromagnetic (EM) devices. With ground truth (empirical evidence) data being generated from a finite-element analysis solver, a deep convolutional neural network is trained in a supervised manner to learn a mapping for magnetic field distribution for topologies of different complexities of geometry, material, and excitation, including a simple coil, a transformer, and a permanent magnet motor. Preliminary experiments show DL model predictions in close agreement with the ground truth. A probabilistic model is introduced to improve the accuracy and to quantify the uncertainty in the prediction, based on Monte Carlo dropout. This paper establishes a basis for a fast and generalizable data-driven model used in the analysis, design, and optimization of EM devices. [ABSTRACT FROM AUTHOR]

Plant species identification represents an extraordinary challenge for machine learning due to visual interspecies similarities and large intraspecies variations. Furthermore, research literature reports that plant species identification usually lacks sufficiently large datasets for training classification models. In this paper, we address this problem with a model that simplifies the VGG-16 architecture, the N-VGG model. The idea behind N-VGG is to reduce experimentally observed overfitting on VGG-16 by using as few trainable parameters as possible. To do this, we substitute the flattening layer on the VGG architecture with a global average pooling layer. This reduces the size of the feature vector. In addition, we eliminate one of the two fully-connected layers and use a new hyper-parameter, N, to indicate the number of nodes on the remaining layer. To show the robustness of the N-VGG model, we conducted extensive experimentation. We trained N-VGG on five datasets for plant species identification. Four of these datasets are publicly available and have been widely used as benchmarks for plant identification models. For all datasets, we compare the accuracy of N-VGG to that of the VGG-16, Inception-v4, and EfficienNet-B3 models. The experimental results show that the N-VGG model achieved the best classification performance for all but one datasets, whereas all the models showed a remarkable performance for the remaining dataset. This evidence supports our initial idea that, for plant species classification, some accuracy might be lost due to overfitting and that having fewer trainable parameters helps in producing a more robust model. [ABSTRACT FROM AUTHOR]

As deep learning models are usually massive and complex, distributed learning is essential for increasing training efficiency. Moreover, in many real-world application scenarios like healthcare, distributed learning can also keep the data local and protect privacy. Recently, the asynchronous decentralized parallel stochastic gradient descent (ADPSGD) algorithm has been proposed and demonstrated to be an efficient and practical strategy where there is no central server, so that each computing node only communicates with its neighbors. Although no raw data will be transmitted across different local nodes, there is still a risk of information leak during the communication process for malicious participants to make attacks. In this paper, we present a differentially private version of asynchronous decentralized parallel SGD framework, or A(DP) $^2$ 2 SGD for short, which maintains communication efficiency of ADPSGD and prevents the inference from malicious participants. Specifically, Rényi differential privacy is used to provide tighter privacy analysis for our composite Gaussian mechanisms while the convergence rate is consistent with the non-private version. Theoretical analysis shows A(DP) $^2$ 2 SGD also converges at the optimal $\mathcal {O}(1/\sqrt{T})$ O (1 / T) rate as SGD. Empirically, A(DP) $^2$ 2 SGD achieves comparable model accuracy as the differentially private version of Synchronous SGD (SSGD) but runs much faster than SSGD in heterogeneous computing environments. [ABSTRACT FROM AUTHOR]

Time series classification models have been garnering significant importance in the research community. However, not much research has been done on generating adversarial samples for these models. These adversarial samples can become a security concern. In this paper, we propose utilizing an adversarial transformation network (ATN) on a distilled model to attack various time series classification models. The proposed attack on the classification model utilizes a distilled model as a surrogate that mimics the behavior of the attacked classical time series classification models. Our proposed methodology is applied onto 1-nearest neighbor dynamic time warping (1-NN DTW) and a fully convolutional network (FCN), all of which are trained on 42 University of California Riverside (UCR) datasets. In this paper, we show both models were susceptible to attacks on all 42 datasets. When compared to Fast Gradient Sign Method, the proposed attack generates a larger faction of successful adversarial black-box attacks. A simple defense mechanism is successfully devised to reduce the fraction of successful adversarial samples. Finally, we recommend future researchers that develop time series classification models to incorporating adversarial data samples into their training data sets to improve resilience on adversarial samples. [ABSTRACT FROM AUTHOR]

Histopathological tissue classification is a simpler way to achieve semantic segmentation for the whole slide images, which can alleviate the requirement of pixel-level dense annotations. Existing works mostly leverage the popular CNN classification backbones in computer vision to achieve histopathological tissue classification. In this paper, we propose a super lightweight plug-and-play module, named Pyramidal Deep-Broad Learning (PDBL), for any well-trained classification backbone to improve the classification performance without a re-training burden. For each patch, we construct a multi-resolution image pyramid to obtain the pyramidal contextual information. For each level in the pyramid, we extract the multi-scale deep-broad features by our proposed Deep-Broad block (DB-block). We equip PDBL in three popular classification backbones, ShuffLeNetV2, EfficientNetb0, and ResNet50 to evaluate the effectiveness and efficiency of our proposed module on two datasets (Kather Multiclass Dataset and the LC25000 Dataset). Experimental results demonstrate the proposed PDBL can steadily improve the tissue-level classification performance for any CNN backbones, especially for the lightweight models when given a small among of training samples (less than 10%). It greatly saves the computational resources and annotation efforts. The source code is available at: https://github.com/linjiatai/PDBL. [ABSTRACT FROM AUTHOR]

Deep learning (DL) techniques have shown promising performance for designing data-driven power system transient stability assessment (TSA) models. However, due to the deep structure of the DL, the resulting model is always a black-box and hard to explain, which hinders its practical adoption by the industry. This paper proposes an interpretable DL-based TSA model to balance the TSA accuracy and transparency. The proposed method combines the strong nonlinear modelling capability of a deep neural network and the interpretability of a decision tree (DT). Through regularizing DL-based model with the average DT path length in the training process, the proposed interpretable DL-based TSA method can visually explain the TSA decision-making process. Simulation results have shown that the proposed method can deliver highly accurate TSA results and interpretable TSA decision-making rules, which can be used for designing preventive control actions. [ABSTRACT FROM AUTHOR]

Automatic target recognition (ATR) is an important part for many computer vision applications. Despite the extensive research which has been carried out in this area for many years, there is no ATR system which performs well on all applications. Recently, different object recognition frameworks have been proposed which yield a high performance in baseline databases. However, our experiments showed that they can fail in real-world scenarios, when dealing with a limited number of data samples. In this paper, we propose a new ATR system, based on deep convolutional neural network (DCNN), to detect the targets in forward looking infrared (FLIR) scenes and recognize their classes. In our proposed ATR framework, a fully convolutional network is trained to map the input FLIR imagery data to a fixed stride correspondingly-sized target score map. The potential targets are identified by applying a threshold on the target score map. Finally, the corresponding regions centered at these target points are fed to a DCNN to classify them into different target types while at the same time rejecting the false alarms. The proposed architecture achieves a significantly better performance in comparison with that of the state-of-the-art methods on two large FLIR image databases. [ABSTRACT FROM AUTHOR]

Deep learning is currently widely used in a variety of applications, including computer vision and natural language processing. End devices, such as smartphones and Internet-of-Things sensors, are generating data that need to be analyzed in real time using deep learning or used to train deep learning models. However, deep learning inference and training require substantial computation resources to run quickly. Edge computing, where a fine mesh of compute nodes are placed close to end devices, is a viable way to meet the high computation and low-latency requirements of deep learning on edge devices and also provides additional benefits in terms of privacy, bandwidth efficiency, and scalability. This paper aims to provide a comprehensive review of the current state of the art at the intersection of deep learning and edge computing. Specifically, it will provide an overview of applications where deep learning is used at the network edge, discuss various approaches for quickly executing deep learning inference across a combination of end devices, edge servers, and the cloud, and describe the methods for training deep learning models across multiple edge devices. It will also discuss open challenges in terms of systems performance, network technologies and management, benchmarks, and privacy. The reader will take away the following concepts from this paper: understanding scenarios where deep learning at the network edge can be useful, understanding common techniques for speeding up deep learning inference and performing distributed training on edge devices, and understanding recent trends and opportunities. [ABSTRACT FROM AUTHOR]

Federated learning (FL) is a collaborative machine learning paradigm, which enables deep learning model training over a large volume of decentralized data residing in mobile devices without accessing clients’ private data. Driven by the ever increasing demand for model training of mobile applications or devices, a vast majority of FL tasks are implemented over wireless fading channels. Due to the time-varying nature of wireless channels, however, random delay occurs in both the uplink and downlink transmissions of FL. How to analyze the overall time consumption of a wireless FL task, or more specifically, a FL’s delay distribution, becomes a challenging but important open problem, especially for delay-sensitive model training. In this paper, we present a unified framework to calculate the approximate delay distributions of FL over arbitrary fading channels. Specifically, saddle point approximation, extreme value theory (EVT), and large deviation theory (LDT) are jointly exploited to find the approximate delay distribution along with its tail distribution, which characterizes the quality-of-service of a wireless FL system. Simulation results will demonstrate that our approximation method achieves a small approximation error, which vanishes with the increase of training accuracy. [ABSTRACT FROM AUTHOR]

Deep learning has aroused a lot of attention and has been used successfully in many domains, such as accurate image recognition and medical diagnosis. Generally, the training of models requires large, representative datasets, which may be collected from a large number of users and contain sensitive information (e.g., users’ photos and medical information). The collected data would be stored and computed by service providers (SPs) or delegated to an untrusted cloud. The users can neither control how it will be used, nor realize what will be learned from it, which make the privacy issues prominent and severe. To solve the privacy issues, one of the most popular approaches is to encrypt users’ data with their public keys. However, this technique inevitably leads to another challenge that how to train the model based on multi-key encrypted data. In this paper, we propose a novel privacy-preserving deep learning model, namely PDLM, to apply deep learning over the encrypted data under multiple keys. In PDLM, lots of users contribute their encrypted data to SP to learn a specific model. We adopt an effective privacy-preserving calculation toolkit to achieve the training process based on stochastic gradient descent (SGD) in a privacy-preserving manner. We also prove that our PDLM can achieve users’ privacy preservation and analyze the efficiency of PDLM in theory. Finally, we conduct an experiment to evaluate PDLM over two real-world datasets and empirical results demonstrate that our PDLM can effectively and efficiently train the model in a privacy-preserving way. [ABSTRACT FROM AUTHOR]

Traditional Deep Neural Network (DNN) security is mostly related to the well-known adversarial input example attack. Recently, another dimension of adversarial attack, namely, attack on DNN weight parameters, has been shown to be very powerful. As a representative one, the Bit-Flip-based adversarial weight Attack (BFA) injects an extremely small amount of faults into weight parameters to hijack the executing DNN function. Prior works of BFA focus on un-targeted attack that can hack all inputs into a random output class by flipping a very small number of weight bits stored in computer memory. This paper proposes the first work of targeted BFA based (T-BFA) adversarial weight attack on DNNs, which can intentionally mislead selected inputs to a target output class. The objective is achieved by identifying the weight bits that are highly associated with classification of a targeted output through a class-dependent vulnerable weight bit searching algorithm. Our proposed T-BFA performance is successfully demonstrated on multiple DNN architectures for image classification tasks. For example, by merely flipping 27 out of 88 million weight bits of ResNet-18, our T-BFA can misclassify all the images from ’Hen’ class into ’Goose’ class (i.e., 100% attack success rate) in ImageNet dataset, while maintaining 59.35% validation accuracy. Moreover, we successfully demonstrate our T-BFA attack in a real computer prototype system running DNN computation, with Ivy Bridge-based Intel i7 CPU and 8GB DDR3 memory. [ABSTRACT FROM AUTHOR]

Click-through rate (CTR) prediction, whose goal is to predict the probability of the user to click on an item, has become increasingly significant in the recommender systems. Recently, some deep learning models with the ability to automatically extract the user interest from his/her behaviors have achieved great success. In these work, the attention mechanism is used to select the user interested items in historical behaviors, improving the performance of the CTR predictor. Normally, these attentive modules can be jointly trained with the base predictor by using gradient descents. In this paper, we regard user interest modeling as a feature selection problem, which we call user interest selection. For such a problem, we propose a novel approach under the framework of the wrapper method, which is named Meta-Wrapper. More specifically, we use a differentiable module as our wrapping operator and then recast its learning problem as a continuous bilevel optimization. Moreover, we use a meta-learning algorithm to solve the optimization and theoretically prove its convergence. Meanwhile, we also provide theoretical analysis to show that our proposed method 1) efficiencies the wrapper-based feature selection, and 2) achieves better resistance to overfitting. Finally, extensive experiments on three public datasets manifest the superiority of our method in boosting the performance of CTR prediction. [ABSTRACT FROM AUTHOR]

Hyperdimensional computing (HDC) has emerged as an alternative lightweight learning solution to deep neural networks. A key characteristic of HDC is the great extent of parallelism that can facilitate hardware acceleration. However, previous hardware implementations of HDC seldom focus on GPU designs, which were also inefficient partly due to the complexity of accelerating HDC on GPUs. In this paper, we present OpenHD, a flexible and high-performance GPU-powered framework for automating the mapping of general HDC applications including classification and clustering to GPUs. OpenHD takes advantage of memory optimization strategies specialized for HDC, minimizing the access time to different memory subsystems, and removing redundant operations. We also propose a novel training method to enable data parallelism in HDC training. Our evaluation result shows that the proposed training rapidly achieves the target accuracy, reducing the required training epochs by 4×. With OpenHD, users can deploy GPU-accelerated HDC applications without domain expert knowledge. Compared to the state-of-the-art GPU-powered HDC implementation, our evaluation on NVIDIA Jetson TX2 shows that OpenHD is up to 10.5× and 314× faster for HDC-based classification and clustering, respectively. Compared with non-HDC classification and clustering on GPUs, OpenHD-based HDC is 11.7× and 53× faster at comparable accuracy. OpenHD is available at: https://github.com/UCSD-SEELab/openhd. [ABSTRACT FROM AUTHOR]

Change/motion detection is a challenging problem in video analysis and surveillance system. Recently, the state-of-the-art methods using the sample-based background model have demonstrated astonishing results with this problem. However, they are ineffective in the dynamic scenes that contain complex motion patterns. In this paper, we introduce a novel data-driven approach that combines the sample-based background model with a feature extractor obtained by training a triplet network. We construct the network by three identical convolutional neural networks, each of which is called a motion feature network. Our network can automatically learn motion patterns from small image patches and transform input images of any size into feature embeddings for high-level representations. The sample-based background model of each pixel is then employed by using the color information and the extracted feature embeddings. We also propose an approach to generate triplet examples from CDNet 2014 for training our network model from scratch. The offline trained network can be used on the fly without re-training on any video sequence before each execution. Therefore, it is feasible for real-time surveillance systems. In this paper, we show that our method outperforms the other state-of-the-art methods on CDNet 2014 and other benchmarks (BMC and Wallflower). [ABSTRACT FROM AUTHOR]

In this paper, the problem of human pose estimation in a 2D still image is addressed. A framework called adaptive holons representation (ADORE) that takes advantage of local and global cues is proposed to improve the pose estimation accuracy. In particular, ADORE is made up of two components: 1) the holons part, independent losses pose nets (ILPNs) is designed to first infer joints location on the global level; and 2) the adaptive part, convolutional local detectors (CLDs) is proposed to subsequently detect the joints in the potential regions generated by ILPN. Pose estimation is formulated as a classification problem toward body joints in ILPN, which consists of two independent loss layers that, respectively, instruct the learning of $x$ and $y$ coordinates of a joint. Experimental results on two challenging benchmark tasks demonstrate that our proposed framework is more efficient than other deep models and has desirable performance. [ABSTRACT FROM AUTHOR]

Convolutional Neural Networks (ConvNets/CNNs) have revolutionized the research in computer vision, due to their ability to capture complex patterns, resulting in high inference accuracies. However, the increasingly complex nature of these neural networks means that they are particularly suited for server computers with powerful GPUs. We envision that deep learning applications will be eventually widely deployed on mobile devices, e.g., smartphones, self-driving cars, and drones. Therefore, in this paper, we aim to understand the resource requirements of CNNs on mobile devices in terms of compute time, memory, and power. First, by deploying several popular CNNs on different mobile CPUs and GPUs, we measure and analyze the performance and resource usage for the CNNs on a layerwise granularity. Our findings point out the potential ways of optimizing the CNN pipelines on mobile devices. Second, we model resource requirements of core computations of CNNs. Finally, based on the measurement and modeling, we build and evaluate our modeling tool, Augur, which takes a CNN configuration (descriptor) as the input and estimates the compute time, memory, and power requirements of the CNN, to give insights about whether and how efficiently a CNN can be run on a given mobile platform. [ABSTRACT FROM AUTHOR]

Deep Learning and Machine Learning models have recently been shown to be effective in many real world applications. While these models achieve increasingly better predictive performance, their structures have also become much more complex. A common and difficult problem for complex models is overfitting. Regularization is used to penalize the complexity of the model in order to avoid overfitting. However, in most learning frameworks, regularization function is usually set with some hyper-parameters where the best setting is difficult to find. In this paper, we propose an adaptive regularization method, as part of a large end-to-end healthcare data analytics software stack, which effectively addresses the above difficulty. First, we propose a general adaptive regularization method based on Gaussian Mixture (GM) to learn the best regularization function according to the observed parameters. Second, we develop an effective update algorithm which integrates Expectation Maximization (EM) with Stochastic Gradient Descent (SGD). Third, we design a lazy update and sparse update algorithm to reduce the computational cost by 4x and 20x, respectively. The overall regularization framework is fast, adaptive, and easy-to-use. We validate the effectiveness of our regularization method through an extensive experimental study over 14 standard benchmark datasets and three kinds of deep learning/machine learning models. The results illustrate that our proposed adaptive regularization method achieves significant improvement over state-of-the-art regularization methods. [ABSTRACT FROM AUTHOR]

In order to retain more feature information of local areas on a point cloud, local grouping and subsampling are the necessary data structuring steps in most hierarchical deep learning models. Due to the disorder nature of the points in a point cloud, the significant time cost may be consumed when grouping and subsampling the points, which consequently results in poor scalability. This paper proposes a fast data structuring method called PSNet (Point Structuring Net). PSNet transforms the spatial features of the points and matches them to the features of local areas in a point cloud. PSNet achieves grouping and sampling at the same time while the existing methods process sampling and grouping in two separate steps (such as using FPS plus kNN). PSNet performs feature transformation pointwise while the existing methods uses the spatial relationship among the points as the reference for grouping. Thanks to these features, PSNet has two important advantages: 1) the grouping and sampling results obtained by PSNet is stable and permutation invariant; and 2) PSNet can be easily parallelized. PSNet can replace the data structuring methods in the mainstream point cloud deep learning models in a plug-and-play manner. We have conducted extensive experiments. The results show that PSNet can improve the training and inference speed significantly while maintaining the model accuracy. [ABSTRACT FROM AUTHOR]

With substantial amount of time, resources and human (team) efforts invested to explore and develop successful deep neural networks (DNN), there emerges an urgent need to protect these inventions from being illegally copied, redistributed, or abused without respecting the intellectual properties of legitimate owners. Following recent progresses along this line, we investigate a number of watermark-based DNN ownership verification methods in the face of ambiguity attacks, which aim to cast doubts on the ownership verification by forging counterfeit watermarks. It is shown that ambiguity attacks pose serious threats to existing DNN watermarking methods. As remedies to the above-mentioned loophole, this paper proposes novel passport-based DNN ownership verification schemes which are both robust to network modifications and resilient to ambiguity attacks. The gist of embedding digital passports is to design and train DNN models in a way such that, the DNN inference performance of an original task will be significantly deteriorated due to forged passports. In other words, genuine passports are not only verified by looking for the predefined signatures, but also reasserted by the unyielding DNN model inference performances. Extensive experimental results justify the effectiveness of the proposed passport-based DNN ownership verification schemes. Code is available at https://github.com/kamwoh/DeepIPR. [ABSTRACT FROM AUTHOR]

Vision-based intelligent driving assistance systems and transportation systems can be improved by enhancing the visibility of the scenes captured in extremely challenging conditions. In particular, many low-image image enhancement (LIE) algorithms have been proposed to facilitate such applications in low-light conditions. While deep learning-based methods have achieved substantial success in this field, most of them require paired training data, which is difficult to be collected. This paper advocates a novel Unsupervised Decomposition and Correction Network (UDCN) for LIE without depending on paired data for training. Inspired by the Retinex model, our method first decomposes images into illumination and reflectance components with an image decomposition network (IDN). Then, the decomposed illumination is processed by an illumination correction network (ICN) and fused with the reflectance to generate a primary enhanced result. In contrast with fully supervised learning approaches, UDCN is an unsupervised one which is trained only with low-light images and corresponding histogram equalized (HE) counterparts (can be derived from the low-light image itself) as input. Both the decomposition and correction networks are optimized under the guidance of hybrid no-reference quality-aware losses and inter-consistency constraints between the low-light image and its HE counterpart. In addition, we also utilize an unsupervised noise removal network (NRN) to remove the noise previously hidden in the darkness for further improving the primary result. Qualitative and quantitative comparison results are reported to demonstrate the efficacy of UDCN and its superiority over several representative alternatives in the literature. The results and code will be made public available at https://github.com/myd945/UDCN. [ABSTRACT FROM AUTHOR]

This study proposes a simple but strong baseline for deep person re-identification (ReID). Deep person ReID has achieved great progress and high performance in recent years. However, many state-of-the-art methods design complex network structures and concatenate multi-branch features. In the literature, some effective training tricks briefly appear in several papers or source codes. The present study collects and evaluates these effective training tricks in person ReID. By combining these tricks, the model achieves 94.5% rank-1 and 85.9% mean average precision on Market1501 with only using the global features of ResNet50. The performance surpasses all existing global- and part-based baselines in person ReID. We propose a novel neck structure named as batch normalization neck (BNNeck). BNNeck adds a batch normalization layer after global pooling layer to separate metric and classification losses into two different feature spaces because we observe they are inconsistent in one embedding space. Extended experiments show that BNNeck can boost the baseline, and our baseline can improve the performance of existing state-of-the-art methods. Our codes and models are available at: https://github.com/michuanhaohao/reid-strong-baseline [ABSTRACT FROM AUTHOR]

Domain adaptation is to transfer the shared knowledge learned from the source domain to a new environment, i.e., target domain. One common practice is to train the model on both labeled source-domain data and unlabeled target-domain data. Yet the learned models are usually biased due to the strong supervision of the source domain. Most researchers adopt the early-stopping strategy to prevent over-fitting, but when to stop training remains a challenging problem since the lack of the target-domain validation set. In this paper, we propose one efficient bootstrapping method, called Adaboost Student, explicitly learning complementary models during training and liberating users from empirical early stopping. Adaboost Student combines deep model learning with the conventional training strategy, i.e., adaptive boosting, and enables interactions between learned models and the data sampler. We adopt one adaptive data sampler to progressively facilitate learning on hard samples and aggregate “weak” models to prevent over-fitting. Extensive experiments show that (1) Without the need to worry about the stopping time, AdaBoost Student provides one robust solution by efficient complementary model learning during training. (2) AdaBoost Student is orthogonal to most domain adaptation methods, which can be combined with existing approaches to further improve the state-of-the-art performance. We have achieved competitive results on three widely-used scene segmentation domain adaptation benchmarks. [ABSTRACT FROM AUTHOR]

We consider the problem of obtaining image quality representations in a self-supervised manner. We use prediction of distortion type and degree as an auxiliary task to learn features from an unlabeled image dataset containing a mixture of synthetic and realistic distortions. We then train a deep Convolutional Neural Network (CNN) using a contrastive pairwise objective to solve the auxiliary problem. We refer to the proposed training framework and resulting deep IQA model as the CONTRastive Image QUality Evaluator (CONTRIQUE). During evaluation, the CNN weights are frozen and a linear regressor maps the learned representations to quality scores in a No-Reference (NR) setting. We show through extensive experiments that CONTRIQUE achieves competitive performance when compared to state-of-the-art NR image quality models, even without any additional fine-tuning of the CNN backbone. The learned representations are highly robust and generalize well across images afflicted by either synthetic or authentic distortions. Our results suggest that powerful quality representations with perceptual relevance can be obtained without requiring large labeled subjective image quality datasets. The implementations used in this paper are available at https://github.com/pavancm/CONTRIQUE. [ABSTRACT FROM AUTHOR]

Over the past years, the increasingly severe data island problem has spawned an emerging distributed deep learning framework—federated learning, in which the global model can be constructed over multiple participants without directly sharing their raw data. Despite its promising prospect, there are still many security challenges in federated learning, such as privacy preservation and integrity verification. Furthermore, federated learning is usually performed with the assistance of a center, which is prone to cause trust worries and communicational bottlenecks. To tackle these challenges, in this paper, we propose a privacy-preserving and verifiable decentralized federated learning framework, named PVD-FL, which can achieve secure deep learning model training under a decentralized architecture. Specifically, we first design an efficient and verifiable cipher-based matrix multiplication (EVCM) algorithm to execute the most basic calculation in deep learning. Then, by employing EVCM, we design a suite of decentralized algorithms to construct the PVD-FL framework, which ensures the confidentiality of both global model and local update and the verification of every training step. Detailed security analysis shows that PVD-FL can well protect privacy against various inference attacks and guarantee training integrity. In addition, the extensive experiments on real-world datasets also demonstrate that PVD-FL can achieve lossless accuracy and practical performance. [ABSTRACT FROM AUTHOR]

Model compression is a widely-used approach for reducing the size of deep learning models without much accuracy loss, enabling resource-hungry models to be compressed for use on resource-constrained devices. In this paper, we study the risk that model compression could provide an opportunity for adversaries to inject stealthy backdoors. In a backdoor attack on a machine learning model, an adversary produces a model that performs well on normal inputs but outputs targeted misclassifications on inputs containing a small trigger pattern. We design stealthy backdoor attacks such that the full-sized model released by adversaries appears to be free from backdoors (even when tested using state-of-the-art techniques), but when the model is compressed it exhibits a highly effective backdoor. We show this can be done for two common model compression techniques—model pruning and model quantization—even in settings where the adversary has limited knowledge of how the particular compression will be done. Our findings demonstrate the importance of performing security tests on the models that will actually be deployed not in their precompressed version. Our implementation is available at https://github.com/yulongtzzz/Stealthy-Backdoors-as-Compression-Artifacts. [ABSTRACT FROM AUTHOR]

Deep learning-based auto-driving systems are vulnerable to adversarial examples attacks which may result in wrong decision making and accidents. An adversarial example can fool the well trained neural networks by adding barely imperceptible perturbations to clean data. In this paper, we explore the mechanism of adversarial examples and adversarial robustness from the perspective of statistical mechanics, and propose an statistical mechanics-based interpretation model of adversarial robustness. The state transition caused by adversarial training based on the theory of fluctuation dissipation disequilibrium in statistical mechanics is formally constructed. Besides, we fully study the adversarial example attacks and training process on system robustness, including the influence of different training processes on network robustness. Our work is helpful to understand and explain the adversarial examples problems and improve the robustness of deep learning-based auto-driving systems. [ABSTRACT FROM AUTHOR]

The ability to predict, anticipate and reason about future outcomes is a key component of intelligent decision-making systems. In light of the success of deep learning in computer vision, deep-learning-based video prediction emerged as a promising research direction. Defined as a self-supervised learning task, video prediction represents a suitable framework for representation learning, as it demonstrated potential capabilities for extracting meaningful representations of the underlying patterns in natural videos. Motivated by the increasing interest in this task, we provide a review on the deep learning methods for prediction in video sequences. We first define the video prediction fundamentals, as well as mandatory background concepts and the most used datasets. Next, we carefully analyze existing video prediction models organized according to a proposed taxonomy, highlighting their contributions and their significance in the field. The summary of the datasets and methods is accompanied with experimental results that facilitate the assessment of the state of the art on a quantitative basis. The paper is summarized by drawing some general conclusions, identifying open research challenges and by pointing out future research directions. [ABSTRACT FROM AUTHOR]

Recently, convolutional neural network (CNN) based models have shown great potential in the task of single image super-resolution (SISR). However, many state-of-the-art SISR solutions are reproducing some tricks proven effective in other vision tasks, such as pursuing a deeper model. In this paper, we propose a new solution (named as Multi-Receptive-Field Network - MRFN), which outperforms existing SISR solutions in three different aspects. First, from receptive field: a novel multi-receptive-field (MRF) module is proposed to extract and fuse features in different receptive fields from local to global. Integrating these hierarchical features can generate better mappings on recovering high-fidelity details at different scales. Second, from network architectures: both dense skip connections and deep supervision are utilized to combine features from the current MRF module and preceding ones for training more representative features. Moreover, a deconvolution layer is embedded at the end of the network to avoid artificial priors induced by numerical data pre-processing (e.g., bicubic stretching), and speed up the restoration process. Finally, from error modeling: different from $L1$ and $L2$ loss functions, we proposed a novel two-parameter training loss called Weighted Huber loss function which can adaptively adjust the value of back-propagated derivative according to the residual value, thus fit the reconstruction error more effectively. Extensive qualitative and quantitative evaluation results on benchmark datasets demonstrate that our proposed MRFN can achieve more accurate recovering results than most state-of-the-art methods with significantly less complexity. [ABSTRACT FROM AUTHOR]

Memory effects of radio frequency power amplifiers (PAs) can interact with dynamic transmitting signals, dynamic operations, and dynamic environment, resulting in complicated nonlinear problems of the PAs. Recently, deep learning based schemes have been proposed to deal with the memory effects. Although these schemes are powerful in constructing complex nonlinear structures, they are still direct learning-based and are relatively static. In this paper, we propose an adaptive deep learning aided digital predistortion (DL-DPD) model by optimizing a deep regression neural network. Thanks to the sequence structure of the proposed DL-DPD, we then make the linearization architecture more adaptive by using multiple sub-DPD modules and an ensemble predicting process. The results show the effectiveness of the proposed adaptive DL-DPD, and reveals that the online system handovers the sub-DPD modules more frequently than expected. [ABSTRACT FROM AUTHOR]

This paper aims at accelerating and compressing deep neural networks to deploy CNN models into small devices like mobile phones or embedded gadgets. We focus on filter level pruning, i.e., the whole filter will be discarded if it is less important. An effective and unified framework, ThiNet (stands for "Thin Net"), is proposed in this paper. We formally establish filter pruning as an optimization problem, and reveal that we need to prune filters based on statistics computed from its next layer, not the current layer, which differentiates ThiNet from existing methods. We also propose "gcos" (Group COnvolution with Shuffling), a more accurate group convolution scheme, to further reduce the pruned model size. Experimental results demonstrate the effectiveness of our method, which has advanced the state-of-the-art. Moreover, we show that the original VGG-16 model can be compressed into a very small model (ThiNet-Tiny) with only 2.66 MB model size, but still preserve AlexNet level accuracy. This small model is evaluated on several benchmarks with different vision tasks (e.g., classification, detection, segmentation), and shows excellent generalization ability. [ABSTRACT FROM AUTHOR]

In the era of digitized images, the goal is to extract information from them and create new knowledge thanks to Computer Vision techniques, Machine Learning and Deep Learning. This enables the use of images for early diagnosis and subsequent treatment of a wide range of diseases. In the dermatological field, deep neural networks are used to distinguish between melanoma and non-melanoma images. In this paper, we have underlined two essential points of melanoma detection research. The first aspect considered is how even a simple modification of the parameters in the dataset determines a change of the accuracy of classifiers. In this case, we investigated the Transfer Learning issues. Following the results of this first analysis, we suggest that continuous training-test iterations are needed to provide robust prediction models. The second point is the need to have a more flexible system architecture that can handle changes in the training datasets. In this context, we proposed the development and implementation of a hybrid architecture based on Cloud, Fog and Edge Computing to provide a Melanoma Detection service based on clinical and dermoscopic images. At the same time, this architecture must deal with the amount of data to be analyzed by reducing the running time of the continuous retrain. This fact has been highlighted with experiments carried out on a single machine and different distribution systems, showing how a distributed approach guarantees output achievement in a much more sufficient time. [ABSTRACT FROM AUTHOR]

With the availability of various sensors in the smartphone, identifying a locomotion mode becomes convenient and effortless in recent years. Information about locomotion mode helps to improve journey planning, travel time estimation, and traffic management. Though there exists a significant amount of work towards locomotion mode recognition, the performance of these work is not pertinent and heavily depends on the labeled training instances. As it is impractical to gather a prior information (labeled instances) about all types of locomotion modes, the recognition model should be able to identify a new or unseen locomotion mode without having any corresponding training instance. This paper proposes a sensors based deep learning model to identify a locomotion mode by using labeled training instances. The approach also incorporates a concept of Zero-Shot learning to identify an unseen locomotion mode. The model obtains an attribute matrix based on the fusion of three semantic matrices. It also constructs a feature matrix by extracting the deep learning and hand-crafted features from the training instances. Later, the model builds a classifier by learning a mapping between attribute and feature matrices. Finally, this work evaluates the performance of the approach on collected and existing datasets using accuracy and F1 score. [ABSTRACT FROM AUTHOR]

In general, an image attribute is a human-nameable visual property that has a semantic connotation. Appropriate modeling of the intrinsic contextual correlations among attributes plays a fundamental role in attribute detection. In this paper, we consider image attribute detection from the perspective of regularized deep learning. In particular, we propose a regularized deep belief network (rDBN) to perform the image attribute detection task, which is composed of two parts: 1) a detection DBN (dDBN) that models the joint distribution of images and their corresponding attributes, which acts as an attribute detector and 2) a contextual restricted Boltzmann machine that explicitly models the correlations among attributes acting as a regularizer that restraints the output detection result given by the dDBN to meet the contextual prior of attributes. Furthermore, we propose an efficient fine-tuning scheme that can further optimize the performance of the dDBN by backpropagation. Experimental results show that the proposed rDBN obtains improvements over the state-of-the-art methods for attribute detection on the benchmark data sets. [ABSTRACT FROM PUBLISHER]

Recently, there has been growing attention on deep learning-based low-light image enhancement algorithms. With this interest, various synthetic low-light image datasets have been released publicly. However, real-world low-light and well-exposed image pair datasets are still lacking. In this paper, we propose a real-world low-light image dataset and a practical lightweight low-light image enhancement network. In order to construct a large-scale real-world low-light dataset, we have not only captured under-exposed images by ourselves but also collected under-exposed images from the Internet. Then, we produce pseudo well-exposed images for each low-light image. Using pairs of a real-world low-light image and a pseudo well-exposed image, we present a lightweight deep CNN model through knowledge distillation. Experimental results demonstrate the effectiveness and practicality of the proposed method on various datasets. [ABSTRACT FROM AUTHOR]

Federated Learning (FL) is a privacy-preserving collaborative learning approach that trains artificial intelligence (AI) models without revealing local datasets of the FL workers. While FL ensures the privacy of the FL workers, its performance is limited by several bottlenecks, which become significant given the increasing amounts of data generated and the size of the FL network. One of the main challenges is the straggler effects where the significant computation delays are caused by the slow FL workers. As such, Coded Federated Learning (CFL), which leverages coding techniques to introduce redundant computations to the FL server, has been proposed to reduce the computation latency. In CFL, the FL server helps to compute a subset of the partial gradients based on the composite parity data and aggregates the computed partial gradients with those received from the FL workers. In order to implement the coding schemes over the FL network, incentive mechanisms are important to allocate the resources of the FL workers and data owners efficiently in order to complete the CFL training tasks. In this paper, we consider a two-level incentive mechanism design problem. In the lower level, the data owners are allowed to support the FL training tasks of the FL workers by contributing their data. To model the dynamics of the selection of FL workers by the data owners, an evolutionary game is adopted to achieve an equilibrium solution. In the upper level, a deep learning based auction is proposed to model the competition among the model owners. [ABSTRACT FROM AUTHOR]

The increasing success of Machine Learning (ML) and Deep Learning (DL) has recently re-sparked interest towards traffic classification. While supervised techniques provide satisfactory performance when classifying known traffic, the detection of zero-day (i.e., unknown) traffic is a more challenging task. At the same time, zero-day detection, generally tackled with unsupervised techniques such as clustering, received less coverage by the traffic classification literature which focuses more on deriving DL models via supervised techniques. Moreover, the combination of supervised and unsupervised techniques poses challenges not fully covered by the traffic classification literature. In this paper, we share our experience on a commercial-grade DL traffic classification engine that combines supervised and unsupervised techniques to identify known and zero-day traffic. In particular, we rely on a dataset with hundreds of very fine grained application labels, and perform a thorough assessment of two state of the art traffic classifiers in commercial-grade settings. This pushes the boundaries of traffic classifiers evaluation beyond the few tens of classes typically used in the literature. Our main contribution is the design and evaluation of GradBP, a novel technique for zero-day applications detection. Based on gradient backpropagation and tailored for DL models, GradBP yields superior performance with respect to state of the art alternatives, in both accuracy and computational cost. Overall, while ML and DL models are both equally able to provide excellent performance for the classification of known traffic, the non-linear feature extraction process of DL models backbone provides sizable advantages for the detection of unknown classes over classical ML models. [ABSTRACT FROM AUTHOR]

Radar radio source (RRS) recognition plays an important role in the fields of military electronic support systems (ESM) and civilian autonomous driving. The rapid development of machine learning technology, especially deep learning, has effectively and efficiently improved RRS intelligent recognition performances when operating in the increasingly complex electromagnetic environment. However, the data sampling limitation and computation cost are still severe challenges in real RRS recognition scenarios. In this paper, we propose a novel network based on meta-transfer learning, called RRSARNet, to achieve effective adaptive RRS recognition in the context of low signal-to-noise ratio (SNR). First, by using the short-time Fourier transform, a six-type small samples RRS simulation dataset with different SNR levels is constructed. Then, a novel RRSARNet, based on metric learning, is proposed, which consists of a four-layer embedding module and a four-layer relational module. Finally, the RRS dataset is divided into training, supporting and testing subsets, which are used to train and test the RRSARNet in a meta-transfer learning method. Experiments on the RRS dataset show that the proposed RRSARNet can achieve an overall accuracy (OA) above 96% and 99% when the SNR is above −15 dB and −10 dB, respectively. Even when the SNR is -30 dB, OA can reach more than 70%. For 5-way 1-shot and 5-way 5-shot experiments, the inference time of an image is about 0.043 and 0.140 milliseconds, respectively. Besides, experiments on the RRS simulation dataset and the two benchmark datasets, the RRSARNet performs better or more competitive than many existing state-of-the-art technologies in terms of recognition accuracy. [ABSTRACT FROM AUTHOR]

A number of visual tracking methods achieve the state-of-the-art performance based on deep learning recently. However, most of these trackers utilize the deep neural network in regression task or classification task separately. In this paper, we propose an adversarial deep tracking framework. The framework is composed of a fully convolutional Siamese neural network (regression network) and a discriminative classification network. Then, we jointly optimize the regression network and the classification network by adversarial learning. In the uniform framework, the regression network and classification network can be trained end-to-end as a whole using large amounts of video training data sets. During the testing phase, the regression network generates a response map which reflects the location and the size of the target within each candidate search patch, and the classification network discriminates which response map is the best in terms of the corresponding template patch and candidate search patch. In addition, we propose an attention visualization algorithm for our tracker, and it reflects the area that attracts the attention of our tracker during tracking. The experimental results on three large-scale visual tracking benchmarks (OTB-100, TC-128, and VOT2016) demonstrate the effectiveness of the proposed tracking algorithm and show that our tracker performs comparably against the state-of-the-art trackers. [ABSTRACT FROM AUTHOR]

In this paper, a novel feature extraction method is proposed for hyperspectral image classification using a Gaussian–Bernoulli restricted Boltzmann machine (GBRBM) in parallel. The proposed approach employs several GBRBMs with different hidden layers to extract deep features from hyperspectral images, which are nonlinear and local invariant. Based on the learned deep features, a logistic regression layer is trained for classification. The proposed approaches are carried out on two public hyperspectral datasets: Pavia University dataset and Salinas dataset, and a new dataset obtained by HySpex imaging spectrometer in the mining area in Xuzhou. The obtained results reveal that the proposed approach offers superior performance compared to traditional classifiers. The advantage of the proposed GBRBM is that it can extract deep features in an unsupervised way and reduce the prediction time by using GPU. In particular, the classification results of the mining area provide valuable suggestions to improve environmental protection. [ABSTRACT FROM AUTHOR]

Deep computation models achieve super performance for big data feature learning. However, training a deep computation model poses a significant challenge since a deep computation model typically involves a large number of parameters. Specially, it needs a high-performance computing server with a large-scale memory and a powerful computing unit to train a deep computation model, making it difficult to increase the size of a deep computation model further for big data feature learning on low-end devices such as conventional desktops and portable CPUs. In this paper, we propose an improved deep computation model based on the canonical polyadic decomposition scheme to compress the parameters and to improve the training efficiency. Furthermore, we devise a learning algorithm based on the back-propagation strategy to train the parameters of the proposed model. The learning algorithm can be directly performed on the compressed parameters to improve the training efficiency. Finally, we carry on the experiments on three representative datasets, i.e., CUAVE, SNAE2, and STL-10, to evaluate the performance of the proposed model by comparing with the conventional deep computation model and other two improved deep computation models based on the Tucker decomposition and the tensor-train network. Results demonstrate that the proposed model can compress parameters greatly and improve the training efficiency significantly with a low classification accuracy drop. [ABSTRACT FROM AUTHOR]

Deep reinforcement learning (DRL) enables autonomic optical networking by allowing the network control and management systems to self-learn successful networking policies from operational experiences. This paper proposes a transfer learning approach for effective and scalable DRL in optical networks. We first present a modular DRL agent design to facilitate retrieving and transferring relevant knowledge between tasks requiring different dimensions of network state data. In particular, we partition network state data into common states, which contain generic information critical to multiple tasks (e.g., the spectrum utilization on fiber links), and task-specific states that are only used by a specific task (e.g., the utilization of virtual network functions). Separate neural network blocks are employed to process different state data. Based on the modular agent design, a multi-task learning (MTL) aided knowledge transferring scheme is proposed. The proposed scheme trains an MTL agent that can master multiple tasks simultaneously and thus enables to learn and transfer better-generalized knowledge across tasks. We perform case studies on the proposed transfer DRL approach considering two scenarios, namely, knowledge transferring between routing, modulation and spectrum assignment (RMSA) tasks for different networks and knowledge transferring from RMSA tasks to anycast service provisioning tasks. The DRL designs for RMSA and anycast service provisioning, including the state, action, and reward formulations and the training mechanisms, are also elaborated. Performance evaluations under both scenarios show that the proposed approach can effectively expedite the training processes of the target tasks and improve the ultimate service throughput. [ABSTRACT FROM AUTHOR]

Performing deep neural network (DNN) inference in real time requires excessive network resources, which poses a big challenge to the resource-limited industrial Internet of things (IIoT) networks. To address the challenge, in this paper, we introduce an end-edge-cloud orchestration architecture, in which the inference task assignment and DNN model placement are flexibly coordinated. Specifically, the DNN models, trained and pre-stored in the cloud, are properly placed at the end and edge to perform DNN inference. To achieve efficient DNN inference, a multi-dimensional resource management problem is formulated to maximize the average inference accuracy while satisfying the strict delay requirements of inference tasks. Due to the mix-integer decision variables, it is difficult to solve the formulated problem directly. Thus, we transform the formulated problem into a Markov decision process which can be solved efficiently. Furthermore, a deep reinforcement learning based resource management scheme is proposed to make real-time optimal resource allocation decisions. Simulation results are provided to demonstrate that the proposed scheme can efficiently allocate the available spectrum, caching, and computing resources, and improve average inference accuracy by 31.4 % compared with the deep deterministic policy gradient benchmark. [ABSTRACT FROM AUTHOR]

In this paper, we propose a structure-coherent deep feature learning method for face alignment. Unlike most existing face alignment methods which overlook the facial structure cues, we explicitly exploit the relation among facial landmarks to make the detector robust to hard cases such as occlusion and large pose. Specifically, we leverage a landmark-graph relational network to enforce the structural relationships among landmarks. We consider the facial landmarks as structural graph nodes and carefully design the neighborhood to passing features among the most related nodes. Our method dynamically adapts the weights of node neighborhood to eliminate distracted information from noisy nodes, such as occluded landmark point. Moreover, different from most previous works which only tend to penalize the landmarks absolute position during the training, we propose a relative location loss to enhance the information of relative location of landmarks. This relative location supervision further regularizes the facial structure. Our approach considers the interactions among facial landmarks and can be easily implemented on top of any convolutional backbone to boost the performance. Extensive experiments on three popular benchmarks, including WFLW, COFW and 300W, demonstrate the effectiveness of the proposed method. In particular, due to explicit structure modeling, our approach is especially robust to challenging cases resulting in impressive low failure rate on COFW and WFLW datasets. The model and code are publicly available at https://github.com/BeierZhu/Sturcture-Coherency-Face-Alignment [ABSTRACT FROM AUTHOR]

The existing fusion-based RGB-D salient object detection methods usually adopt the bistream structure to strike a balance in the fusion trade-off between RGB and depth (D). While the D quality usually varies among the scenes, the state-of-the-art bistream approaches are depth-quality-unaware, resulting in substantial difficulties in achieving complementary fusion status between RGB and D and leading to poor fusion results for low-quality D. Thus, this paper attempts to integrate a novel depth-quality-aware subnet into the classic bistream structure in order to assess the depth quality prior to conducting the selective RGB-D fusion. Compared to the SOTA bistream methods, the major advantage of our method is its ability to lessen the importance of the low-quality, no-contribution, or even negative-contribution D regions during RGB-D fusion, achieving a much improved complementary status between RGB and D. Our source code and data are available online at https://github.com/qdu1995/DQSD. [ABSTRACT FROM AUTHOR]

While convolutional neural network (CNN) has achieved overwhelming success in various vision tasks, its heavy computational cost and storage overhead limit the practical use on mobile or embedded devices. Recently, compressing CNN models has attracted considerable attention, where pruning CNN filters, also known as the channel pruning, has generated great research popularity due to its high compression rate. In this paper, a new channel pruning framework is proposed, which can significantly reduce the computational complexity while maintaining sufficient model accuracy. Unlike most existing approaches that seek to-be-pruned filters layer by layer, we argue that choosing appropriate layers for pruning is more crucial, which can result in more complexity reduction but less performance drop. To this end, we utilize a long short-term memory (LSTM) to learn the hierarchical characteristics of a network and generate a global network pruning scheme. On top of it, we propose a data-dependent soft pruning method, dubbed Squeeze-Excitation-Pruning (SEP), which does not physically prune any filters but selectively excludes some kernels involved in calculating forward and backward propagations depending on the pruning scheme. Compared with the hard pruning, our soft pruning can better retain the capacity and knowledge of the baseline model. Experimental results demonstrate that our approach still achieves comparable accuracy even when reducing 70.1% Floating-point operation per second (FLOPs) for VGG and 47.5% for Resnet-56. [ABSTRACT FROM AUTHOR]

Recent years have witnessed the success of deep neural networks in dealing with a plenty of practical problems. Dropout has played an essential role in many successful deep neural networks, by inducing regularization in the model training. In this paper, we present a new regularized training approach: Shakeout. Instead of randomly discarding units as Dropout does at the training stage, Shakeout randomly chooses to enhance or reverse each unit's contribution to the next layer. This minor modification of Dropout has the statistical trait: the regularizer induced by Shakeout adaptively combines L0 , L1 and L2 regularization terms. Our classification experiments with representative deep architectures on image datasets MNIST, CIFAR-10 and ImageNet show that Shakeout deals with over-fitting effectively and outperforms Dropout. We empirically demonstrate that Shakeout leads to sparser weights under both unsupervised and supervised settings. Shakeout also leads to the grouping effect of the input units in a layer. Considering the weights in reflecting the importance of connections, Shakeout is superior to Dropout, which is valuable for the deep model compression. Moreover, we demonstrate that Shakeout can effectively reduce the instability of the training process of the deep architecture. [ABSTRACT FROM AUTHOR]