Search

As network complexity increases because of the diversification of applications and services based on internet of autonomous things, it is difficult for humans to design the optimal control rule for software-defined networking controllers. Intelligence-defined networking, called IDN, is proposed to overcome this limitation through machine learning algorithms. Since the existing IDN approaches are mostly designed to optimize only the network quality of services, including throughput, jitter, and latency, the controllers don’t consider the importance of data in the applications and the services. This causes the controller to allocate insufficient resources to the crucial data flow, which leads critical problems, such as self-driving car accidents. To prevent this problem, we propose an importance-based IDN architecture that enables network controllers to manage network traffic with importance levels of data flows. Firstly, we devise an importance estimation scheme to set the importance level for the flows. Secondly, a dynamic resource allocation model of the controllers is developed by means of deep learning algorithms in order to make the optimal network resource. Additionally, an online learning mechanism based on weighted auto-labeling is adopted to continue enhancing the adaptability of the resource allocation model on runtime as the network conditions change. The evaluation results of the proposed architecture under various autonomous things scenarios show that the loss rate for data flows of higher importance is reduced by one-quarter compared to the case of a network controller without importance level and that the bandwidth waste ratio is reduced by ten percent compared to the rule-based model.

Long short-term memory (LSTM) applications need fast yet compact models. Neural network compression approaches have been promising for cutting down network complexity by skipping insignificant weights. However, current strategies remain hardware-agnostic and network complexity reduction does not always translate to execution efficiency. We propose a hardware-guided symbiotic training methodology for compact, accurate, yet execution-efficient inference models. It is based on our observation that hardware may introduce substantial non-monotonic behavior, which we call the latency hysteresis effect, when evaluating network size vs. latency. Leveraging hardware-impacted hysteresis effect and sparsity, we enable a symbiosis of model compactness and accuracy with execution efficiency. We have evaluated our approach on language modeling and speech recognition applications. Relative to the traditional LSTM architecture obtained for the Penn Treebank dataset, we reduce the number of parameters by 18.0x (30.5x) and run-time latency by up to 2.4x (5.2x) on Nvidia GPUs (Intel CPUs) without any accuracy degradation. For the DeepSpeech2 architecture obtained for the AN4 dataset, we reduce the model size by 7.0x (19.4x), word error rate from 12.9% to 9.9% (10.4%), and run-time latency by up to 1.7x (2.4x) on Nvidia GPUs (Intel CPUs). Our method consistently outperforms prior art, yielding compact, accurate, yet execution-efficient models.

It has been widely recognized that the efficient training of neural networks (NNs) is crucial to classification performance. While a series of gradient-based approaches have been extensively developed, they are criticized for the ease of trapping into local optima and sensitivity to hyperparameters. Due to the high robustness and wide applicability, evolutionary algorithms (EAs) have been regarded as a promising alternative for training NNs in recent years. However, EAs suffer from the curse of dimensionality and are inefficient in training deep NNs (DNNs). By inheriting the advantages of both the gradient-based approaches and EAs, this article proposes a gradient-guided evolutionary approach to train DNNs. The proposed approach suggests a novel genetic operator to optimize the weights in the search space, where the search direction is determined by the gradient of weights. Moreover, the network sparsity is considered in the proposed approach, which highly reduces the network complexity and alleviates overfitting. Experimental results on single-layer NNs, deep-layer NNs, recurrent NNs, and convolutional NNs (CNNs) demonstrate the effectiveness of the proposed approach. In short, this work not only introduces a novel approach for training DNNs but also enhances the performance of EAs in solving large-scale optimization problems.

The theory and experiment of quantum information have been studied extensively in recent years, and the feasibility of quantum communication has been proved. Although the fundamental technology is not yet mature, research on quantum internet should be conducted. To implement quantum internet, an architecture that describes how quantum nodes are linked to form networks and how protocol functions are vertically composed need to be developed urgently. In this paper, we present a novel design of a cluster-based structure to describe how quantum nodes are interconnected, and how the structure can improve the performance of qubit transmission and reduce the network complexity. The idea of the quantum local area network (QLAN) is proposed as an essential component of the quantum internet. Besides, each quantum repeater links to neighboring repeaters to form a core network, and multiple QLANs are connected through the core network. The core network can be grouped into different hierarchical quantum repeater networks according to needed service requirements. For the sake of interoperability and fast prototyping, we adopt the idea of OSI layering model of the current Internet in the design of quantum internet. Finally, we elaborate on the composition of quantum nodes and the realization of end-to-end communication.

Recently, using deep learning(DL) in super-resolution(SR) has ac- hieved great success. These methods combine the convolutional neural network(CNN) to learn a general matrix function for an end-to-end mapping. However, as the width and depth of the network increase, there are two essential problems in the SR tasks. On the one hand, a wider and deeper network will bring better performance but increase the calculating complexity and memory consumption. On the other hand, the expanded architecture will miss the intermediate feature details in the information transmitting process. Hence, a SISR(Single Image Super-Resolution) network with wider feature information blocks(WFIB) is proposed to address these issues by making a balance between the network complexity and performance. Cascade residual block(CRB) helps the network make full use of contextual feature information. Extensive experiments verify that our network achieves better performance and with fewer parameters than the state-of-the-art methods.

During the last decade, deep neural networks have shown a great performance in many machine learning tasks such as classification and clustering. One of the most successful networks is the CNN (Convolutional Neural Network), which has been applied in many application domains such as pattern recognition, medical diagnosis, and signal processing. Despite the very interesting performance of CNNs, their architecture design is still so far a major challenge for researchers and practitioners. Several works have been proposed in the literature with the aim to find optimized architectures such as ResNet and VGGNet. Unfortunately, most of these architectures are either manually defined by experts or automatically designed by greedy induction algorithms. Recent works suggest the use of Evolutionary Algorithms (EAs) thanks to their ability to escape locally-optimal architectures. Despite the fact that EAs have shown interesting performance, researchers in this direction have considered the design task as a single-level optimization problem; which represents the main research gap we tackle in this paper. The main contribution behind our work consists in the fact that CNN architecture design has a hierarchical nature and thus could be seen as a Bi-Level Optimization Problem (BLOP) where: (1) the upper level minimizes the network complexity defined by the number of blocks and the number of nodes per block; and (2) the lower level optimizes the convolution block ‘graphs’ topologies by maximizing the classification accuracy. Motivated by the originality of our observation with respect to the state of the art, we frame for the first time the CNN architecture design problem as a BLOP and then solve it using an adapted version of an existing efficient bi-level EA; through the definition of the solution encoding, the fitness function, and the variation operators at each level. The adapted EA is named BLOP-CNN and is assessed on the image classification task using the commonly employed CIFAR-10 and CIFAR-100 benchmark data sets. The analysis of our experimental results show the merits of our proposed method in providing the user with optimized architectures that outperform many recent and prominent architectures coming from the three different approaches, namely: manual design, reinforcement learning-based generation, and evolutionary optimization. Moreover, to show the applicability of our approach, we have conducted a case study on the detection of the COVID-19 using a set of benchmark chest X-ray and Computed Tomography (CT) images.

Over the last decade, wireless devices have developed rapidly until predictions will develop with high complexity and dynamic. So that new capabilities are needed for wireless problems in this problem. Software Defined Network (SDN) is generally a wire-based network, but to meet the needs of users in terms of its implementation, it has begun to introduce a Wireless-based SDN called Software Defined Wireless Network (SDWN) which provides good service quality and reach and higher tools, so as to be able to provide new capabilities to wireless in a high complexity and very dynamic. When SDN is implemented in a wireless network it will require a routing solution that chooses paths due to network complexity. In this paper, SDWN is tested by being applied to mesh topologies of 4,6 and 8 access points (AP) because this topology is very often used in wireless-based networks. To improve network performance, Dijkstra's algorithm is added with the user mobility scheme used is RandomDirection. The Dijkstra algorithm was chosen because it is very effective compared to other algorithms. The performance measured in this study is Quality of Service (QoS), which is a parameter that indicates the quality of data packets in a network. The measurement results obtained show that the QoS value in this study meets the parameters considered by the ITU-T G1010 with a delay value of 1.3 ms for data services and packet loss below 0.1%. When compared with the ITU-T standard, the delay and packet loss fall into the very good category.

Background Liver cancer is the sixth most common cancer worldwide. It is mostly diagnosed with a computed tomography scan. Nowadays deep learning methods have been used for the segmentation of the liver and its tumor from the computed tomography (CT) scan images. This research mainly focused on segmenting liver and tumor from the abdominal CT scan images using a deep learning method and minimizing the effort and time used for a liver cancer diagnosis. The algorithm is based on the original UNet architecture. But, here in this paper, the numbers of filters on each convolutional block were reduced and new batch normalization and a dropout layer were added after each convolutional block of the contracting path. Results Using this algorithm a dice score of 0.96, 0.74, and 0.63 were obtained for liver segmentation, segmentation of tumors from the liver, and the segmentation of tumor from abdominal CT scan images respectively. The segmentation results of liver and tumor from the liver showed an improvement of 0.01 and 0.11 respectively from other works. Conclusion This work proposed a liver and a tumor segmentation method using a UNet architecture as a baseline. Modification regarding the number of filters and network layers were done on the original UNet model to reduce the network complexity and improve segmentation performance. A new class balancing method is also introduced to minimize the class imbalance problem. Through these, the algorithm attained better segmentation results and showed good improvement. However, it faced difficulty in segmenting small and irregular tumors.

Narrowband Internet of Things (NB-IoT) has been utilized in a vast range of applications. The conventional NB-IoT is subjected to the limitation of reduced coverage, limited power, and delay. The conventional NB-IoT uses orthogonal frequency-division multiplexing with Long Term Evolution (LTE) is applied, which increases the network complexity. The increase in network complexity increases power consumption and delay. To overcome those limitations, modified Reduced Fast Fourier Transmission (MRFFT) sample generation technique is used in every Uplink transmission for in CAT-0 and CAT-4 based LTE implementation technique. Multiple sample frequency transmission techniques will be required for Up Link (UL) based on control information for getting from Down Link (DL). In this paper, an effective R2MDC—MRFFT technique has been used to combine hardware and data flow complexity. The proposed MRFFT architecture uses “Feedback (R2)-Commutator (R4) MRFFT” is utilized for reducing latency and hardware complexity of the data path flow part of NB-IoT in LTE. The design of feedback-commutator architecture provides the same functionality with increasing throughput (2.4 GSamples) with 30.605 ns latency reduction, lower power consumption, and reduced coverage enhancement. The proposed feedback-commutator based frequency transmission technique is compared with the traditional frequency transmission technique used in CAT-0, CAT-4 devices. This enhanced MRFFT sample generation technique is incorporated into data flow heterogeneous architecture of NB-IoT design procedure to improve the design performances of LTE wireless communication in the perspective of coverage in low mobility.

Air-writing recognition has received wide attention due to its potential application in intelligent systems. To date, some of the fundamental problems in isolated writing have not been addressed effectively. This paper presents a simple yet effective air-writing recognition approach based on deep convolutional neural networks (CNNs). A robust and efficient hand tracking algorithm is proposed to extract air-writing trajectories collected by a single web camera. The algorithm addresses the push-to-write problem and avoids restrictions on the users’ writing without using a delimiter and an imaginary box. A novel preprocessing scheme is also presented to convert the writing trajectory into appropriate forms of data, making the CNNs trained with these forms of data simpler and more effective. Experimental results indicate that the proposed approach not only obtains much higher recognition accuracy but also reduces the network complexity significantly compared to the popular image-based methods.

Traditional tracking algorithms reply on manually extracted features while deep learning algorithms can automatically extract features, which is of great importance for single target tracking in the complex underwater environment that is characterized by poor visibility, low contrasts, and occlusion. The convolutional neural network SiamRPN++ is an advanced deep learning algorithm whose backbone network structure is. However, this algorithm is difficult to be implemented in an underwater platform due to its large amount of computation and high requirement for the computing power of the hardware. To solve this problem, a new backbone network structure NewNet-62 characterized by inverted residual bottleneck block and the SiamRPN++ algorithm are proposed. The depth-wise separable convolution in NewNet-62 simplifies the computation so that the SiamRPN++ algorithm can realize fast and accurate tracking of underwater single targets. The video set in VOT2016 was used to test the algorithm. The results show that the SiamRPN++ algorithm using NewNet-62 as the main network has the great performance. Compared with the algorithms using ResNet-50, the accuracy of the SiamRPN++ algorithm increased to 0.524 and its EAO increased to 0.303. And its tracking speed increased to 73.74. The network complexity reduced to 3.013 billion and the network parameters reduced to 12.538 million, which significantly reduces network complexity and number of network parameters.

In recent years, few-shot learning is proposed to solve the problem of lacking samples in deep learning. However, previous works are mainly concentrated on optimizing neural network structures or augmenting the dataset while ignoring the local relationship of the images. Considering that humans pay more attention to the foreground or prominent features of the images during image recognition, we proposed the foreground feature attention module (FFAM) based on an unsupervised saliency detector for few-shot learning. The FFAM consists of two parts: the foreground extraction module and the features attention module. More specifically, we first extract the foreground images by Robust Background Detector (RBD), one of the best unsupervised saliency detectors. Secondly, we employ the same embedding module to extract the features of both original images and foreground images. Finally, we introduce three improvements to enhance the foreground features and make our network focus on the foreground features without losing background information. Our proposed FFAM is more sensitive to the foreground features than previous approaches. Hence, it effectively recognizes those images with similar backgrounds. Extensive experiments are conducted on miniImagenet and tieredImagenet datasets. It is demonstrated that our proposed FFAM greatly improves the accuracy performance over baseline systems for both one-shot and few-shot classification tasks without increasing the network complexity.

Neural Architecture Search (NAS), i.e., the automation of neural network design, has gained much popularity in recent years with increasingly complex search algorithms being proposed. Yet, solid comparisons with simple baselines are often missing. At the same time, recent retrospective studies have found many new algorithms to be no better than random search (RS). In this work we consider the use of a simple Local Search (LS) algorithm for NAS. We particularly consider a multi-objective NAS formulation, with network accuracy and network complexity as two objectives, as understanding the trade-off between these two objectives is arguably among the most interesting aspects of NAS. The proposed LS algorithm is compared with RS and two evolutionary algorithms (EAs), as these are often heralded as being ideal for multi-objective optimization. To promote reproducibility, we create and release two benchmark datasets, named MacroNAS-C10 and -C100, containing 200K saved network evaluations for two established image classification tasks, CIFAR-10 and CIFAR-100. Our benchmarks are designed to be complementary to existing benchmarks, especially in that they are better suited for multi-objective search. We additionally consider a version of the problem with a much larger architecture space. While we find and show that the considered algorithms explore the search space in fundamentally different ways, we also find that LS substantially outperforms RS and even performs nearly as good as state-of-the-art EAs. We believe that this provides strong evidence that LS is truly a competitive baseline for NAS against which new NAS algorithms should be benchmarked.

Recently, the consolidation of Social Network and the Internet of Things (IoT) acquiring appreciation and consideration among researchers due to its immense and adaptable behavior. Social Internet of Things (SIoT) comprehends the adequacy to support the inventive applications and network services with adept and persuasive demeanors. The author’s exertion for this research provides the subsistence to the SIoT amelioration Currently, the SIoT technique proven to be competent, but the exponential growth of heterogeneous smart devices may arise a challenging scenario. Searching the right object, service, or best path to the informative nodes in such highly convoluted networks is always a challenging task. This research strengthen the SIoT paradigm by consolidating the properties of the individual’s Small World. In addition, this research also advises a Smart Social Agent (SSA) to minimize the human intervention. The absolute objective of this paper is two-fold, (i) transformation of SIoT network into Individual’s Small World SIoT network, (ii) searching the right information efficiently by exploiting SSA. To accomplish the goal we have proposed an efficient Query based Searching algorithm which exploits Service Oriented properties of Individual’s Small World SIoT network and searches information. First result shows the significant alleviation in the network complexity, reduces the average path length, and network diameter. Eventually, it improves the network scalability and navigability. While second result shows that the proposed Query based searching algorithm proves to be efficient with its mean execution time not more than 2.66 ms for complete network on two different machines. Whereas, for Individual’s small world network the mean execution time is not more than 2.47 ms for both machine.

For the detection of marine ship objects in radar images, large-scale networks based on deep learning are difficult to be deployed on existing radar-equipped devices. This paper proposes a lightweight convolutional neural network, LiraNet, which combines the idea of dense connections, residual connections and group convolution, including stem blocks and extractor modules. The designed stem block uses a series of small convolutions to extract the input image features, and the extractor network adopts the designed two-way dense connection module, which further reduces the network operation complexity. Mounting LiraNet on the object detection framework Darknet, this paper proposes Lira-you only look once (Lira-YOLO), a lightweight model for ship detection in radar images, which can easily be deployed on the mobile devices. Lira-YOLO's prediction module uses a two-layer YOLO prediction layer and adds a residual module for better feature delivery. At the same time, in order to fully verify the performance of the model, mini-RD, a lightweight distance Doppler domain radar images dataset, is constructed. Experiments show that the network complexity of Lira-YOLO is low, being only 2.980 Bflops, and the parameter quantity is smaller, which is only 4.3 MB. The mean average precision (mAP) indicators on the mini-RD and SAR ship detection dataset (SSDD) reach 83.21% and 85.46%, respectively, which is comparable to the tiny-YOLOv3. Lira-YOLO has achieved a good detection accuracy with less memory and computational cost.

In existing works of deep learning-based resource allocation, the scalability degrades heavily with the increase of network complexity, which is due to their limited learning ability of shallow neural networks and insufficient knowledge of network. Nowadays, to address the growth of cell density, cooperative beamforming in user-centric network (UCN) is emerged, where the additional degrees of freedom of multi-antenna and cell coordination aggravate the challenges. This letter proposes a deep residual learning framework, UcnBeamNet, to enhance the ability of approximating the iterative algorithm for sum rate maximization, where multi-branch subnets are connected in parallel to extract extra information. Specifically, a weighted minimum mean square error (WMMSE)-based algorithm is derived to determine the optimal clusters and beamforming matrices; then UcnBeamNet is trained to learn the input-output mapping and provide direct insight of UCN from association matrices in addition to plural inputs. Extensive experiments demonstrate UcnBeamNet still reaches 90.38% sum-rate relative to conventional algorithm even with a large network size, and achieves more than 50, $000\times $ speed up in computational efficiency.

Software-Defined Networking (SDN) is an innovational network architecture which gives network administrators the ability to directly control the whole network by programming on a centralized controller. Due to network complexity, networks are unlikely to be bug-free. The ability to verify data plane properties will make network management easier for network administrators in SDN. In this paper, we present a novel atomic predicates based data plane properties verification method for SDN using Spark which is a big data processing framework. First, we verify packet reachability which is a fundamental data plane property. Then, we verify other data plane properties such as loop-freedom and nonexistence of black holes. In addition, the proposed method can detect a security threat existing in SDN called firewall bypass threat with packet reachability verification. By adopting atomic predicates, we achieve less computational and storage overhead. We implement the methods and study the performance. The results of experiments show that we can efficiently and accurately detect loops, black holes and firewall bypass threats.

Mobile network operators (MNOs) are in the process of overlaying their conventional macro cellular networks with shorter range cells such as outdoor pico cells. The resultant increase in network complexity creates substantial overhead in terms of operating expenses, time, and labor for their planning and management. Artificial intelligence (AI) offers the potential for MNOs to operate their networks in a more organic and cost-efficient manner. We argue that deploying AI in fifth generation (5G) and beyond will require surmounting significant technical barriers in terms of robustness, performance, and complexity. We outline future research directions, identify top five challenges, and present a possible roadmap to realize the vision of AI-enabled cellular networks for Beyond- 5G and sixth generation (6G) networks.

In this article, we propose a variant of the electric vehicle routing problem considering explicitly the intermediate nodes. Ultimately aiming to provide an optimal routing plan for the shuttle fleet that serves to a university settlement internally, we consider a real road network by explicitly taking into account in the formulation the entire intersections existing and the time-varying passenger demand at shuttle stops, as well as the vehicle dynamics, battery, and recharging features. On purpose, a mathematical program to obtain the joint minimization of a number of objectives in terms of cost, i.e., vehicle operating, battery recharging, and recharging station purchasing, is formulated. Solutions employing an exact method are sought using models of mixed integer program within scenarios involving a number of features including the campus-wide passenger demand, seat capacity of shuttles, battery capacity of shuttles, and recharging station location. Further solutions to our model formulation have been obtained using a benchmark set of instances designed for a large-scale real network. Our findings show that considering a real road network as it is, is significant in exact routing solutions despite the fact that the level of network complexity is an issue. We suggest that a trade-off among the actuality of the network topology and the consequent computational load should be carefully made in order to obtain solutions using exact methods. As is shown, there is room to investigate further in details the dynamics of routing considering especially the effects of changes in flow conditions at intermediate nodes using our formulation.

Urban parks are essential components of urban ecosystems, providing recreation and relaxation places to residents. Measuring the spatial accessibility to urban parks serves as an initial step in urban planning and developing urban development strategies to improve social and environmental justice. This study aims to evaluate measures of spatial accessibility to urban parks by comparing three geographic information systems (GIS)-based approaches, accounting for network complexity, transport modes, distance thresholds, and destination choices. Taking Ipswich City (Australia) and Enschede (the Netherlands) as two testbeds, we examine the spatial patterns of a total of 21 accessibility measures in the two cities and conduct a correlation and principal component analysis to unravel the interrelationship between these measures. The results suggest that among all measures under the three approaches, the selection of distance thresholds and transport modes matter more to accessibility measures than the destination choices. Furthermore, when distance threshold and transport mode are held constant, the network-based and entrance-based methods provide more realistic accessibility measures than other methods. We also discuss the generality of the entrance-based method we propose and suggest ways to choose the most appropriate accessibility measure for use in different contexts.

The COVID-19 pandemic has disrupted supply chain operations globally. Nevertheless, resilient firms have the capacity to combat an unprecedented situation with the right strategic approach. The current research has developed an integrated research model that combines factors such as supply chain intelligence, supply chain communication, leadership commitment, risk management orientation, supply chain capability and network complexity to investigate supply chain resilience. The research model of this study was empirically tested with 309 responses collected from supply chain managers. Results revealed that supply chain resilience is measured with supply chain intelligence, supply chain communication, leadership commitment, risk management orientation, supply chain capability and network complexity and demonstrated a substantial variance R2 of 0.548% towards supply chain resilience. Practically, this study suggests that supply chain managers should focus on factors such as big data analytics, risk management orientation,1 supply chain communication and leadership commitment to enhance supply chain resilience and sustainable supply chain performance.

The possibility of connecting billions of smart end devices in the Internet of Things (IoT) provides wide range of services to the user. But, the unlimited connectivity of devices in IoT brings security issues when it is connected to wireless networks. Integrating cloud with IoT networks gains more attention as it reduces the sensor node resource limitations. However, the network complexity, open broadcast characteristics of IoT networks are vulnerable to attacks. To ensure network security and reliable operations, Intrusion Detection Systems (IDS) are widely preferred. IDS identifies the anomalies effectively in complex network environments and ensures the security of the network. Traditional intrusion detection systems based on neural networks consume long training time and low classification accuracy. Recently, deep learning methods are widely used in various image and signal processing, security applications. This research work presents a deep learning-based intrusion detection system for multi-cloud IoT environment to overcome the limitations of neural network-based intrusion detection models. The proposed intrusion detection model improves the detection accuracy by improving the training efficiency. Experimental evaluation of proposed model using NSL-KDD dataset provides improved performance than conventional techniques attaining 97.51% of detection rate, 96.28% of detection accuracy, and 94.41% of precision.

In this paper, we have investigated and benchmarked the augmentation approach of image augmentation to increase or provide a different result in detection accuracy compared to the basic method without augmentation. We explored two different methods of pixel-wise operations: pixel domain manipulation and spatial domain transformation to analyze the effect of increasing data for typical defect detection problems. We used two object detection models Faster R-CNN and Cascade R-CNN on top of ResNet-50 as our baseline models. To gain accuracy, we found that the effectiveness of data augmentation for defect detection is influenced by network complexity and the surface defect properties.

Deep learning approaches to estimating full 3D orientations of objects, in addition to object classes, are limited in their accuracies, due to the difficulty in learning the continuous nature of three-axis orientation variations by regression or classification with sufficient generalization. This paper presents a novel progressive deep learning framework, herein referred to as 3D POCO Net, that offers high accuracy in estimating orientations about three rotational axes yet with efficiency in network complexity. The proposed 3D POCO Net is configured, using four PointNet-based networks for independently representing the object class and three individual axes of rotations. The four independent networks are linked by in-between association subnetworks that are trained to progressively map the global features learned by individual networks one after another for fine-tuning the independent networks. In 3D POCO Net, high accuracy is achieved by combining a high precision classification based on a large number of orientation classes with a regression based on a weighted sum of classification outputs, while high efficiency is maintained by a progressive framework by which a large number of orientation classes are grouped into independent networks linked by association subnetworks. We implemented 3D POCO Net for full three-axis orientation variations and trained it with about 146 million orientation variations augmented from the ModelNet10 dataset. The testing results show that we can achieve an orientation regression error of about 2.5° with about 90% accuracy in object classification for general three-axis orientation estimation and object classification. Furthermore, we demonstrate that a pre-trained 3D POCO Net can serve as an orientation representation platform based on which orientations as well as object classes of partial point clouds from occluded objects are learned in the form of transfer learning.

Facial expression emotion recognition is an intuitive reflection of a person’s mental state, which contains rich emotional information, and is one of the most important forms of interpersonal communication. It can be used in various fields, including psychology. As a celebrity in ancient China, Zeng Guofan’s wisdom involves facial emotion recognition techniques. His book Bing Jian summarizes eight methods on how to identify people, especially how to choose the right one, which means “look at the eyes and nose for evil and righteousness, the lips for truth and falsehood; the temperament for success and fame, the spirit for wealth and fortune; the fingers and claws for ideas, the hamstrings for setback; if you want to know his consecution, you can focus on what he has said.” It is said that a person’s personality, mind, goodness, and badness can be showed by his face. However, due to the complexity and variability of human facial expression emotion features, traditional facial expression emotion recognition technology has the disadvantages of insufficient feature extraction and susceptibility to external environmental influences. Therefore, this article proposes a novel feature fusion dual-channel expression recognition algorithm based on machine learning theory and philosophical thinking. Specifically, the feature extracted using convolutional neural network (CNN) ignores the problem of subtle changes in facial expressions. The first path of the proposed algorithm takes the Gabor feature of the ROI area as input. In order to make full use of the detailed features of the active facial expression emotion area, first segment the active facial expression emotion area from the original face image, and use the Gabor transform to extract the emotion features of the area. Focus on the detailed description of the local area. The second path proposes an efficient channel attention network based on depth separable convolution to improve linear bottleneck structure, reduce network complexity, and prevent overfitting by designing an efficient attention module that combines the depth of the feature map with spatial information. It focuses more on extracting important features, improves emotion recognition accuracy, and outperforms the competition on the FER2013 dataset.

The process of product design specification is subjective in nature and this motivates the application of approaches and methods to make it explicit. Among the available approaches, the systematization of the development of a single product, modular products, family of products and products for mass customization stands out. Among the methods, the design structure matrix (DSM) is highlighted, as well as the use of networks to represent the variables and their interdependence relations. Representation is very important to increase the cognitive capacity of those involved in the design and to facilitate communication between specialists and non-specialists. The clarification of the knowledge and reasoning of the design increases the complexity of the specification, which needs to be managed. In this work, the method called design structure network (DSN) is proposed, allowing the visualization of the design variables as nodes of a network and the relations of interdependence as links and the specification reasoning can be represented as a path that connects the nodes in a network. For the management of network complexity, ten principles based on cognitive processes are implemented. The DSN method was applied in the geometric specification of surfboard, and the results obtained show the potential of graphical representation of the specification reasoning, as well as the ability to reduce the complexity of the network.

Overfitting is a common problem for computer vision applications It is a problem that when training convolution neural networks and is caused by lack of training data or network complexity. The novel sequence-dropout (SD) method is proposed in this paper to alleviate the problem of overfitting when training networks. The SD method works by dropping out units (channels of feature) from the network in a sequence, replacing the traditional operation of random omitting. Sophisticated aggregation strategies are used to obtain the global information of feature channels, and channel-wise weights are produced by gating mechanism. The SD method then selectively drops out the feature channels according to the channel-wise weights that represent the importance degree of each channel. The proposed SD block can be plugged into state-of-the-art backbone CNN models such as VGGNet and ResNet. The SD block is then evaluated on these models, demonstrating consistent performance gains over the baseline model on widely-used benchmark image classification datasets including MNIST, CIFAR-10, CIFAR-100, and ImageNet2012. Experimental results demonstrate that the superior performance of the SD block compared to other modern methods.

Skin Lesion detection and classification are very critical in diagnosing skin malignancy. Existing Deep learning-based Computer-aided diagnosis (CAD) methods still perform poorly on challenging skin lesions with complex features such as fuzzy boundaries, artifacts presence, low contrast with the background and, limited training datasets. They also rely heavily on a suitable turning of millions of parameters which often leads to over-fitting, poor generalization, and heavy consumption of computing resources. This study proposes a new framework that performs both segmentation and classification of skin lesions for automated detection of skin cancer. The proposed framework consists of two stages: the first stage leverages on an encoder-decoder Fully Convolutional Network (FCN) to learn the complex and inhomogeneous skin lesion features with the encoder stage learning the coarse appearance and the decoder learning the lesion borders details. Our FCN is designed with the sub-networks connected through a series of skip pathways that incorporate long skip and short-cut connections unlike, the only long skip connections commonly used in the traditional FCN, for residual learning strategy and effective training. The network also integrates the Conditional Random Field (CRF) module which employs a linear combination of Gaussian kernels for its pairwise edge potentials for contour refinement and lesion boundaries localization. The second stage proposes a novel FCN-based DenseNet framework that is composed of dense blocks that are merged and connected via the concatenation strategy and transition layer. The system also employs hyper-parameters optimization techniques to reduce network complexity and improve computing efficiency. This approach encourages feature reuse and thus requires a small number of parameters and effective with limited data. The proposed model was evaluated on publicly available HAM10000 dataset of over 10000 images consisting of 7 different categories of diseases with 98% accuracy, 98.5% recall, and 99% of AUC score respectively.

Deep learning has been successfully utilized in synthetic aperture radar (SAR) automatic target recognition tasks and obtained state-of-the-art results. However, current deep learning algorithms do not perform well when SAR images are occluded, noisy, or with a great depression angle variance. This article proposes a novel method, SAR capsule network, to achieve the accurate and robust classification of SAR images without significantly increasing network complexity. Specifically, we develop a convolutional neural network extension based on Hinton's capsule network to capture spatial relationships specialized in classification between different entities in a SAR image. The SAR capsules are learned by a vector-based full connection operation instead of the traditional routing process, which not only alleviates the computational burden but also improves recognition accuracy. For occlusion, additive noise, and multiplicative noise tests, SAR capsule network shows superior robustness compared with typical convolution neural networks. When missing training data in a certain aspect angle range or existing a large depression angle variance between training data and test data, the proposed network achieves better performance than the existing works and reveals some competitive advantages in several test scenarios.

Driven by the growth of cloud and virtualization technologies, telecom services are under pressure to become increasingly dynamic. Cloud services can be instantiated in a matter of minutes on computing platforms, which can rapidly schedule and allocate virtualized physical resources. A similar flexibility to adapt to dynamic service requests remains a challenge with network resources. Recently, centralized controllers have been deployed in data center networks to adapt network connectivity to computational needs. However, due to the complexity and variety of carrier networking technologies, carrier networks have been slow to move towards centralized control. More recently, carriers are evolving to address network complexity by pushing for the disaggregation of network elements and the introduction of equipment with more control flexibility. Carriers require a new control paradigm to benefit from this new flexibility, which extends far beyond today's OpenFlow network controllers. The lack of a global network controller with a consistent network view slows down the development and deployment of the more dynamic network services required by extensive cloud technologies. To address the scaling requirements of the carrier environment, the envisioned network controller would be implemented as a distributed system. In this article, we describe a global network controller designed to function as a network operating system with multi-domain, multi-layer capabilities that will transform carrier-grade telecom services. We define the requirements for a carrier-grade network operating system (OS) and describe a prototype system (NetUNIX) that can control heterogeneous networks at large scale and ensure high service reliability. We highlight the advantages of the NetUNIX platform in the context of other open source platforms and illustrate the use of the NetUNIX platform with two use cases: a metro-distributed data center and 5G next-generation fronthaul.

Purpose The purpose of this paper is to consider one of the recent and practical extensions of the resource-constrained project scheduling problem (RCPSP) termed as the multi-skill resource-constrained project scheduling problem (MSRCPSP) for investigation. The objective is the minimization of the makespan or total project duration. Design/methodology/approach To solve this complex problem, the authors propose a teaching–learning-based optimization (TLBO) algorithm in which self-study and examination have been used as additional features to enhance its exploration and exploitation capabilities. An activity list-based encoding scheme has been modified to include the resource assignment information because of the multi-skill nature of the algorithm. In addition, a genetic algorithm (GA) is also developed in this work for the purpose of comparisons. The computational experiments are performed on 216 test instances with varying complexity and characteristics generated for the purpose. Findings The results obtained after computations show that the TLBO has performed significantly better than GA in terms of average percentage deviation from the critical path-based lower bound for different combinations of three parameters, namely, skill factor, network complexity and modified resource strength. Research limitations/implications The modified TLBO proposed in this paper can be conveniently applied to any product or service organization wherein human resources are involved in executing project activities. Practical implications The developed model can suitably handle resource allocation problems faced in real-life large-sized projects usually administered in software development companies, consultancy firms, R&D-based organizations, maintenance firms, big construction houses, etc. wherein human resources are involved. Originality/value The current work aims to propose an effective metaheuristic for a more realistic version of MSRCPSP, in which resource requirements of activities may be more than one. Moreover, to enhance the exploration and exploitation capabilities of the original TLBO, the authors use two additional concepts, namely, self-study and examination in the search process.

The human brain is an organic which whole consisting of hundreds of billions of neurons, and the neural network is the information path of the brain for processing information. Network characteristics will change along with network topology or node change, and the change may lead to brain lesions. Using EEG to construct brain network can effectively discover the internal mechanism of improvement in human brain, but the traditional method of constructing brain network is difficult to reflect the network complexity because of the small number of network nodes. In this paper, a new algorithm of brain network construction based on chaotic time series is proposed. The method estimates neuronal dynamic equations and neuronal networks topology by using only noisy time series. By analyzing the network constructed based on phase space reconstruction of Lorenz system, network characteristics that the average path length increase as embedding dimension, and the cluster coefficient on the contrary are found. The topology and statistical characteristics of the network constructed from neuronal chaotic bursting can reflect time-evolution of time series. Through analyzing the real epileptic EEG signals dataset using the proposed complex networks construction method, we found that epileptic seizure is different from seizure-free intervals in network topology and statistical characteristics. Epileptic EEG of before, during and after a tonic-clonic seizure indicate that cluster coefficients of the sliding local networks significantly increased during a seizure, so the result can provide clonic seizure prediction. The proposed method of complex network topology estimation and complex network construction of time series can depict the dynamic characteristics of brain network, and provide a new idea for the treatment of mental illness.

Visual object tracking is a core technology and challenging research in computer vision. With the rapid development of unmanned aerial vehicle (UAV), more and more UAVs are equipped with video cameras to conduct object tracking. Researching object tracking of UAV image sequences is of great help for practical applications. Recent trackers based on deep networks have a great success by learning an effective representation of targets, however, the offline training is time-consuming, and the complex network may reduce its efficiency. In this paper, to effectively decrease the training time and handle the problem of network complexity, an effective tracking algorithm called multi-scale gcForest tracking (MSGCF) is proposed. First, to enrich the sample information and overcome the problem of scale variations, the multi-scale transformation is used for helping network to obtain the stronger representation of target samples. Then, gcForest that is a decision tree ensemble approach is utilized in tracking task to extract the target features, and gcForest is much easier to train and works well in small-scale training data. Most importantly, we make improvements to the gcForest by adding three channels at the top of network to adapt to multi-scale images and improve the representation ability of network. Furthermore, computation speed can be increased greatly through the method of reducing feature dimension based on compressive sensing (CS) theory, and the power of features can also be reserved. Finally, a support vector machine (SVM) classifier is employed to effectively separate targets and backgrounds. Extensive experimental results on UAV image sequences and challenging benchmark data sets demonstrate that the proposed MSGCF algorithm is effective.

Deep learning methods have recently found widespread adoption for remote sensing tasks, particularly in image or pixel classification. Their flexibility and versatility has enabled researchers to propose many different designs to process remote sensing data in all spectral, spatial, and temporal dimensions. In most of the reported cases they surpass their non-deep rivals in overall classification accuracy. However, there is considerable diversity in implementation details in each case and a systematic quantitative comparison to non-deep classifiers does not exist. In this paper, we look at the major research papers that have studied deep learning image classifiers in recent years and undertake a meta-analysis on their performance compared to the most used non-deep rival, Support Vector Machine (SVM) classifiers. We focus on mono-temporal classification as the time-series image classification did not offer sufficient samples. Our work covered 103 manuscripts and included 92 cases that supported direct accuracy comparisons between deep learners and SVMs. Our general findings are the following: (i) Deep networks have better performance than non-deep spectral SVM implementations, with Convolutional Neural Networks (CNNs) performing better than other deep learners. This advantage, however, diminishes when feeding SVM with richer features extracted from data (e.g. spatial filters). (ii) Transfer learning and fine-tuning on pre-trained CNNs are offering promising results over spectral or enhanced SVM, however these pre-trained networks are currently limited to RGB input data, therefore currently lack applicability in multi/hyperspectral data. (iii) There is no strong relationship between network complexity and accuracy gains over SVM; small to medium networks perform similarly to more complex networks. (iv) Contrary to the popular belief, there are numerous cases of high deep networks performance with training proportions of 10% or less. Our study also indicates that the new generation of classifiers is often overperforming existing benchmark datasets, with accuracies surpassing 99%. There is a clear need for new benchmark dataset collections with diverse spectral, spatial and temporal resolutions and coverage that will enable us to study the design generalizations, challenge these new classifiers, and further advance remote sensing science. Our community could also benefit from a coordinated effort to create a large pre-trained network specifically designed for remote sensing images that users could later fine-tune and adjust to their study specifics.

The state‐of‐the‐art video coding standard, Versatile Video Coding (VVC) or H.266, has demonstrated its superior coding efficiency over its predecessor HEVC/H.265. In this paper, a novel in‐loop filter based on convolutional neural network (CNN) is illustrated to further improve the coding efficiency over VVC. In this filter, one single NN model is used to process multiple video components simultaneously. In addition, with a quality map generated for each video component as network input, the same single NN model is capable of processing videos in different qualities and resolutions while maintaining coding efficiency, which reduces the overall network complexity significantly. Simulation results show that the proposed approach provides average BD‐rate savings of 6.27%, 18.78% and 20.42% under AI configuration, and average BD-rate savings of 5.18%, 21.95% and 22.13% under RA configuration, respectively for Y, Cb and Cr components.

Vehicle pose estimation is useful for applications such as self-driving cars, traffic monitoring, and scene analysis. Recent developments in computer vision and deep learning have achieved significant progress in human pose estimation, but little of this work has been applied to vehicle pose. We propose VehiPose, an efficient architecture for vehicle pose estimation, based on a multi-scale deep learning approach that achieves high accuracy vehicle pose estimation while maintaining manageable network complexity and modularity. The VehiPose architecture combines an encoder-decoder architecture with a waterfall atrous convolution module for multi-scale feature representation. Our approach aims to reduce the loss due to successive pooling layers and preserve the multiscale contextual and spatial information in the encoder feature representations. The waterfall module generates multiscale features, as it leverages the efficiency of progressive filtering while maintaining wider fields-of-view through the concatenation of multiple features. This multi-scale approach results in a robust vehicle pose estimation architecture that incorporates contextual information across scales and performs the localization of vehicle keypoints in an end-to-end trainable network.

Crop disease diagnosis is of great significance to crop yield and agricultural production. The image identification methods have become the standard methods to address challenges related to disease diagnosis. However, it is still difficult to diagnose accurately for real-life agricultural environment. It has diverse disease conditions and complex backgrounds. We propose an improved deep convolution neural network to identify tomato leaf diseases better for real-life agriculture environments. The proposed network includes residual blocks, a multiscale feature extraction module for extracting multiscale lesions, and an attention module that captures cross-dimension interaction and extracts lesion features better from complex backgrounds. Extensive experimental results show that the proposed network achieves the average identification accuracy of 88.64% on the dataset with the complex backdrop. It also proves that the method has significant advantages in terms of network complexity and realtime performance compared with other models. Moreover, on the other public dataset, the proposed method achieves brilliant results. It is certified that the attention module recognizes disease spots more accurately and has few parameters relative to another state-of-the-art method. The results demonstrate that the proposed model provides a high-performance solution for crop diagnosis under the real agricultural environment.

Convolutional neural networks (CNNs) achieved great success on various tasks in recent years. Their applications to low power and low cost hardware platforms, however, have been often limited due to extensive complexity of convolution layers. We present a new class of transform domain deep neural networks (DNNs), where convolution operations are replaced by element-wise multiplications in heterogeneous transform domains. To further reduce the network complexity, we propose a framework to learn sparse-orthogonal weights in heterogeneous transform domains co-optimized with a hardware-efficient accelerator architecture to minimize the overhead of handling sparse weights. Furthermore, sparse-orthogonal weights are non-uniformly quantized with canonical-signed-digit (CSD) representations to substitute multiplications with simpler additions. The proposed approach reduces the complexity by a factor of 4.9– 6.8 $\times$ without compromising the DNN accuracy compared to equivalent CNNs that employ sparse (pruned) weights. The code is available at https://github.com/unchenyu/HTNN.

Traditional image classification methods mostly use offline learning mode, which takes a lot of time when data is updated. In this paper, we propose a density-sorting-based convolutional fuzzy min-max neural network (DCFMNN) for image classification to solve this problem. DCFMNN is realized based on convolutional Neural Network (CNN) and density-sorting-based fuzzy min-max neural network. CNN is applied for image feature extraction. Density-sorting-based fuzzy min-max neural network is used for classification, which includes density-based sorting part and fuzzy min-max (FMM) neural network part. In the part of density-based sorting, patterns are sorted according to the points with the highest density in the same class and two densest points are considered for selection. The purpose is to overcome the influence of the pattern input order in the original FMM on the creation of the hyperbox. In the part of FMM, the fuzzy set classification method is used to enable online learning. Diverse CNN architectures are applied to DCFMNN. The benchmark image datasets were employed for evaluation on DCFMNN. Experimental results show that DCFMNN has high classification accuracy and less network complexity, and its online learning ability reduces the training time.

In this paper, two semantic segmentation models, DO-UNet and DO-LinkNet, are presented for the detection of human settlements, and a threshold-based model is proposed to detect areas with electricity. In DO-UNet and DO-LinkNet, the conventional convolutional layer is replaced with depthwise over-parameterized convolutional layer. Also, an extra pooling operation is carried out in the last layer since the size of the input images is different from that of the labels. Depthwise over-parameterized convolutional layer enhances the convolutional layer with an additional depthwise convolution. Pooling operation can accelerate training speed, increase the receptive field in feature extraction, and reduce the requirement of network complexity. In the detection of settlements without electricity challenge track, our best F1-score on the validation set and the test set are 0.8820 and 0.8798, respectively.

Partial occlusion is a common issue encountered in 3D human model acquisition using single viewpoint. The use of deep generative model is recently drawn into point clouds reconstruction to infer full point clouds due to its learning capability. In this paper, the reconstruction of 3D human point clouds from sparse partial point clouds acquired from a single viewpoint depth camera is developed using an autoencoder architecture. The proposed autoencoder architecture is comprised of a novel joints encoder and a patch decoder. The joints encoder first learns the latent representation of input partial point clouds using Gaussian maximum likelihood and infers a set of Gaussian components consisting of the localized skeletal keypoints and its variances of a 3D human model. Subsequently, the inferred Gaussian components are viewed as local human part keypoints for local patch sampling using a patch decoder. The proposed network demonstrated adequate estimated skeletal joints precision and reconstruction fidelity evaluated using a non-synthetic dataset. In addition, the network has less network complexity compared to existing work and able to provide a concise 3D human representations using estimated skeletal joints and its variances.

Crop disease diagnosis is of great significance to crop yield and agricultural production. Deep learning methods have become the main research direction to solve the diagnosis of crop diseases. This paper proposed a deep convolutional neural network that integrates an attention mechanism, which can better adapt to the diagnosis of a variety of tomato leaf diseases. The network structure mainly includes residual blocks and attention extraction modules. The model can accurately extract complex features of various diseases. Extensive comparative experiment results show that the proposed model achieves the average identification accuracy of 96.81% on the tomato leaf diseases dataset. It proves that the model has significant advantages in terms of network complexity and real-time performance compared with other models. Moreover, through the model comparison experiment on the grape leaf diseases public dataset, the proposed model also achieves better results, and the average identification accuracy of 99.24%. It is certified that add the attention module can more accurately extract the complex features of a variety of diseases and has fewer parameters. The proposed model provides a high-performance solution for crop diagnosis under the real agricultural environment.

The 5G technology brings new services such as the faster speeds of enhanced mobile broadband, mass deployment of machine-type communications, and ultra-reliable low-latency services. Consequently, it comes to a critical need to design a network infrastructure that can easily and efficiently support various service. Service Level Agreement (SLA) based services were traditionally established in the Internet Protocol / Multiprotocol Label Switching (IP/MPLS) network by implementing traffic engineering using protocols such as Resource Reservation Protocol - Traffic Engineering (RSVP-TE). However, it is not easy for some network operators to implement the current 5G development as mentioned earlier. Segment Routing (SR) presents as a technology to minimize network complexity and provide a fundamental for supporting SLA services, flexibility, reliability, and overall performance in the core of a network. In this study, we simulate and implement SR into MPLS Layer 3 Virtual Private Network (L3VPN) intra-AS domain to show that (i) reducing any transport protocol will not affect the service performance; instead, it simplifies service provisioning and problem handling (ii) when a network failure happens in the network, using Topology Independent Loop-Free Alternate (TI-LFA) and microloops avoidance can provide better failover time than conventional Fast Re-Route (FRR) (iii) SLA path differentiation for services within the same Virtual Routing and Forwarding (VRF) can be applied using Segment Routing - Traffic Engineering (SR-TE). It will be an essential foundation for network slicing in the future.

Intent-Based Networking (IBN) is a model that enables proactive network control and automation to satisfy high-level demands. It follows a closed-loop mechanism that abstracts the network complexity and dynamically allows updates to the network monitoring and intelligence. Hence, it eliminates the traditionally practiced error-prone manual network control. In contrast with the traditional reactive approach, IBN-based approach promotes the proactive and dynamic solution. To this end, IBN uses machine learning (ML) to predict the future and keeps a balance between the intended and actual state. Hence, this work considers an IBN-driven routing mechanism to ensure service provisioning with the desired quality of service (QoS). In contrast to the traditional static routing scheme, this work considers AI-driven dynamic and proactively updatable routing schemes to ensure QoS. This work demonstrates AI-driven service route control mechanism for ensuring quality of service (QoS) on top of Korea Advanced Research Network-Software Defined Infrastructure (KOREN-SDI). More precisely, an ML-based approach is used to find the best path between source and destination nodes. In addition, it incorporates IBN closed-loop process to monitor and update service routes proactively based on predicted future link utilization. This novel approach introduces proactive updates on runtime to avoid future failures and ensures seamless service provisioning with the fulfillment of QoS demands by changing traffic routes dynamically.

More and more tenants (enterprises and personal users) migrate their tasks to clouds since it is a simple and low-cost way to obtain enough computing resources. However, due to potential node failures and malicious tenants, the modern cloud encounters one critical challenge, i.e., robustness. Conventionally, the cloud vendors deploy auxiliary systems to protect the cloud, which requires additional resource cost and increases the network complexity. To enhance the system robustness, this paper proposes a complementary scheme to improve the cloud robustness through efficient VM placement. Specifically, to alleviate the impact of malicious tenants and node failures on the cloud, when deploying VMs, we limit the number of pods (or service nodes) that each tenant can access, and the number of tenants hosted by each pod (or service node). Though there are a lot of works on VM placement, it is very challenging when the robustness issue is taken into consideration. To solve this problem, we formulate an integer linear programming and propose a rounding-based algorithm with a logarithmic approximation ratio. The simulation results show the high efficiency of the proposed algorithm. For example, our algorithm can improve the network throughput by 150% with other alternatives.

With increasing network complexity, intelligent mechanisms to efficiently achieve the required quality of service of wireless-enabled applications are being developed, especially for industrial environments due to the onset of the fourth industrial revolution. In this paper, the potential benefits of wireless channel quality prediction for two of the three major use cases supported by 5G viz. enhanced Mobile BroadBand (eMBB) and Ultra-Reliable Low Latency Communication (URLLC) are quantified in an industrial indoor environment through simulations. Our analysis shows that the ability to perform perfect prediction improves the 10th user throughput percentile by up to 125% for eMBB use case and decreases the 90th resource utilization percentile by up to 37% for URLLC use case. Furthermore, the maximum tolerable prediction inaccuracy is found to be up to 5 dB and 0.35 dB for eMBB and URLLC use cases, respectively.

Image compression is a method to remove spatial redundancy between adjacent pixels and reconstruct a high-quality image. In the past few years, deep learning has gained huge attention from the research community and produced promising image reconstruction results. Therefore, recent methods focused on developing deeper and more complex networks, which significantly increased network complexity. In this paper, two effective novel blocks are developed: analysis and synthesis block that employs the convolution layer and Generalized Divisive Normalization (GDN) in the variable-rate encoder and decoder side. Our network utilizes a pixel RNN approach for quantization. Furthermore, to improve the whole network, we encode a residual image using LSTM cells to reduce unnecessary information. Experimental results demonstrated that the pro-posed variable-rate framework with novel blocks outperforms existing methods and standard image codecs, such as George’s [11] and JPEG in terms of image similarity. The project page along with code and models are available at https://github.com/khawar512/cvpr_image_compress

Neural Architecture Search (NAS) aims to automatically discover optimal architectures. In this paper, we propose an hourglass-inspired approach (HourNAS) for extremely fast NAS. It is motivated by the fact that the effects of the architecture often proceed from the vital few blocks. Acting like the narrow neck of an hourglass, vital blocks in the guaranteed path from the input to the output of a deep neural network restrict the information flow and influence the network accuracy. The other blocks occupy the major volume of the network and determine the overall network complexity, corresponding to the bulbs of an hourglass. To achieve an extremely fast NAS while preserving the high accuracy, we propose to identify the vital blocks and make them the priority in the architecture search. The search space of those non-vital blocks is further shrunk to only cover the candidates that are affordable under the computational resource constraints. Experimental results on ImageNet show that only using 3 hours (0.1 days) with one GPU, our HourNAS can search an architecture that achieves a 77.0% Top-1 accuracy, which outperforms the state-of-the-art methods.