Search

Pruning methods to compress and accelerate deep convolutional neural networks (CNNs) have recently attracted growing attention, with the view of deploying pruned networks on resource-constrained hardware devices. However, most existing methods focus on small granularities, such as weight, kernel and filter, for the exploration of pruning. Thus, it will be bound to iteratively prune the whole neural networks based on those small granularities for high compression ratio with little performance loss. To address these issues, we theoretically analyze the relationship between the activation and gradient sparsity, and the channel saliency. Based on our findings, we propose a novel and effective method of weak sub-network pruning (WSP). Specifically, for a well-trained network model, we divide the whole compression process into two non-iterative stages. The first stage is to directly obtain a strong sub-network by pruning the weakest one. We first identify the less important channels from all the layers and determine the weakest sub-network, whereby each selected channel makes a minimal contribution to both the feed-forward and feed-backward processes. Then, a one-shot pruning strategy is executed to form a strong sub-network enabling fine tuning, while significantly reducing the impact of the network depth and width on the compression efficiency, especially for deep and wide network architectures. The second stage is to globally fine-tune the strong sub-network using several epochs to restore its original recognition accuracy. Furthermore, our proposed method impacts on the fully-connected layers as well as the convolutional layers for simultaneous compression and acceleration. Comprehensive experiments on VGG16 and ResNet-50 involving a variety of popular benchmarks, such as ImageNet-1K, CIFAR-10, CUB-200 and PASCAL VOC, demonstrate that our WSP method achieves superior performance on classification, domain adaption and object detection tasks with small model size. Our source code is available at https://github.com/QingbeiGuo/WSP.git. [ABSTRACT FROM AUTHOR]

Although group convolution operators are increasingly used in deep convolutional neural networks to improve the computational efficiency and to reduce the number of parameters, most existing methods construct their group convolution architectures by a predefined partitioning of the filters of each convolutional layer into multiple regular filter groups with an equal spatial group size and data-independence, which prevents a full exploitation of their potential. To tackle this issue, we propose a novel method of designing self-grouping convolutional neural networks, called SG-CNN, in which the filters of each convolutional layer group themselves based on the similarity of their importance vectors. Concretely, for each filter, we first evaluate the importance value of their input channels to identify the importance vectors, and then group these vectors by clustering. Using the resulting data-dependent centroids, we prune the less important connections, which implicitly minimizes the accuracy loss of the pruning, thus yielding a set of diverse group convolution filters. Subsequently, we develop two fine-tuning schemes, i.e. (1) both local and global fine-tuning and (2) global only fine-tuning, which experimentally deliver comparable results, to recover the recognition capacity of the pruned network. Comprehensive experiments carried out on the CIFAR-10/100 and ImageNet datasets demonstrate that our self-grouping convolution method adapts to various state-of-the-art CNN architectures, such as ResNet and DenseNet, and delivers superior performance in terms of compression ratio, speedup and recognition accuracy. We demonstrate the ability of SG-CNN to generalize by transfer learning, including domain adaption and object detection, showing competitive results. Our source code is available at https://github.com/QingbeiGuo/SG-CNN.git. [ABSTRACT FROM AUTHOR]

• We build a new standalone control module based on the scaled sigmoid function to enrich the neural network module family to enable the neural architecture optimization. • Our DNAL method produces no candidate neural architectures but one, thus drastically improving the learning efficiency, i.e., costing 20 epochs for CIFAR-10 and 10 epochs for ImageNet. • It is applicable to conventional CNNs, lightweight CNNs, and stochastic supernets. • Extensive experiments confirm that our DNAL method achieves excellent performance on various CNN architectures, including VGG16, ResNet50, MobileNetV2, and ProxylessNAS, over the task of CIFAR-10 and ImageNet-1K classification. Efficient neural networks has received ever-increasing attention with the evolution of convolutional neural networks (CNNs), especially involving their deployment on embedded and mobile platforms. One of the biggest problems to obtaining such efficient neural networks is efficiency, even recent differentiable neural architecture search (DNAS) requires to sample a small number of candidate neural architectures for the selection of the optimal neural architecture. To address this computational efficiency issue, we introduce a novel architecture parameterization based on scaled sigmoid function , and propose a general Differentiable Neural Architecture Learning (DNAL) method to obtain efficient neural networks without the need to evaluate candidate neural networks. Specifically, for stochastic supernets as well as conventional CNNs, we build a new channel-wise module layer with the architecture components controlled by a scaled sigmoid function. We train these neural network models from scratch. The network optimization is decoupled into the weight optimization and the architecture optimization, which avoids the interaction between the two types of parameters and alleviates the vanishing gradient problem. We address the non-convex optimization problem of efficient neural networks by the continuous scaled sigmoid method instead of the common softmax method. Extensive experiments demonstrate our DNAL method delivers superior performance in terms of efficiency, and adapts to conventional CNNs (e.g., VGG16 and ResNet50), lightweight CNNs (e.g., MobileNetV2) and stochastic supernets (e.g., ProxylessNAS). The optimal neural networks learned by DNAL surpass those produced by the state-of-the-art methods on the benchmark CIFAR-10 and ImageNet-1K dataset in accuracy, model size and computational complexity. Our source code is available at https://github.com/QingbeiGuo/DNAL.git. [ABSTRACT FROM AUTHOR]

Image retrieval based on deep learning of hash has made great progress. The hash method increases retrieval speed greatly while saving storage space. However, some problems exist, such as the distinctiveness of image feature still needs to be improved and some image features are still redundant. We propose a new deep learning to hash method, namely, deep multiscale divergence hashing, which provides a high diversity and compact image feature for image retrieval. The discriminative features from deep neural networks are identified by the importance criterion in network pruning techniques and the feature redundancy is reduced. Then, the selected features across different layers are fused in a certain proportion to increase the diversity of features for image retrieval. We also present a hybrid loss function in hash space, which consists of the weighted pairwise cross-entropy loss function and the KL-divergence. It tends to minimize the hamming distance between similar images and maximize the hamming distance between different images, which helps improve the accuracy. Massive experimental results show that our method achieves better feature distinguishability and more advanced image retrieval accuracy, and surpasses HashNet by 11.46%, 7.58%, and 13.86% on MS COCO, NUS-WIDE, and CIFAR-10 datasets, respectively. [ABSTRACT FROM AUTHOR]

During the COVID-19 pandemic, online classes became the only option for many students. The main challenge for these classes was conducting risky and complex chemical or biological experiments in a domestic environment. To address this challenge, a smart experiment system called MRLab was developed. MRLab used wearables such as a smart glove and head-mounted device to record sensory data and a multimodal hybrid fusion model GVVS to interpret the user's experimental intent, which essentially transforms the user's abstract behavioral actions into a probabilistic set of experimental intent that can be computed. Different experiments in MRLab used different libraries of experimental intents. The SrNet model in GVVS was used to estimate the probability of the user's gesture behavior generated from the smart glove, while the SIPA algorithm compared speech information entered during the experiment with the experimental intent library to estimate the probability of the user's intent. At the same time, the scene visual channel monitored the information about the object the user intended to operate, with the SVF algorithm computing the probability of the intended object in real-time. The results from ANOVA and post-hoc comparative testing conducted on 21 volunteers revealed that MRLab outperformed other experiment modes, including WEB, AR, and VR, with a higher intention understanding rate, efficiency, and user satisfaction. Therefore, MRLab proved to be a useful alternative to traditional physics laboratory experiments during the pandemic, along with being an additional teaching tool for remote learning purposes. [ABSTRACT FROM AUTHOR]

Abstract The naturalness and safety of human-computer interaction have always been primary research focuses in the field of human-computer interaction. This paper proposes a multimodal intention understanding algorithm (MIUIC), which incorporates comfort analysis, as a solution to address the issues of low intention understanding rate, weak interaction, and weak collaboration that are often observed in most massage systems. The algorithm efficiently fuses multimodal data based on objective implicit information to address the challenge of low intention understanding rates caused by non-standard expression of natural behavior. Moreover, this algorithm incorporates comfort reasoning to detect and address intentions related to security threats while providing the ability for robots to make behavioral decisions through inverse active interaction, leading to more equitable human-robot interactions. To test the validity and safety of the MIUIC algorithm, we embedded the algorithm into a mechanical arm massage system. Subsequently, 45 elderly volunteers were invited to participate in experimental tests. Finally, to verify the validity and safety of the MIUIC algorithm, we assessed the algorithm in terms of four aspects, including multimodal intention recognition rate, the ability to reduce data dispersion, the intention enhancement rate under reverse human-machine interaction, and the rate of avoiding dangerous intentions. In conclusion, the MIUIC algorithm enhances the intention understanding rate and promotes. [ABSTRACT FROM AUTHOR]

In wireless sensor networks, node localization is the first operation after deployment, and the efficiency of localization can directly influence the performance of the wireless sensor network, so designing an efficient localization algorithm has been an important focus in wireless sensor network research. Currently, anchor-free algorithms are the focus of research in node localization because they have the advantages of flexible structure, low cost, high stability and good extensibility. In this article, related research studies of unique localization-, angle- and distance-obtaining methods are introduced, and existing algorithms are analysed. Based on the analysis of existing research, an anchor-free localization algorithm named ladder diffusion node localization algorithm is proposed, in which sensor nodes are located by directional diffused activation and acknowledgement packets with fixed energy. In ladder diffusion node localization algorithm, the time cost of localization is greatly reduced, the communication tasks in localization are allocated in a more balanced manner, and node locations are obtained mainly by rigorous computation instead of measurement, which can reduce transmission measurement error. Finally, the performance of ladder diffusion node localization algorithm is proved both in theory and by the simulation results of experiments. [ABSTRACT FROM AUTHOR]

Data aggregation is an important method of improving transmitting efficiency of WSNs, but existing researches have some disadvantages as follows: several periods delay will be generated when filtering messages; the reduplicated messages filtering ratio is low; complex calculations need to be executed; and extra hardware should be added to obtain high performance. To resolve these problems, this paper proposes a real time and high efficient data aggregation scheme (dynamical message list based data aggregation, DMLDA) based on clustering routing algorithm. DMLDA includes three procedures: activating nodes, clustering nodes, and filtering messages. In filtering procedure, a special data structure named dynamic list will be established in every filtering node; it is designed to store messages transmitted by filtering node, and, comparing current message with list’s items, the message’s redundancy can be judged without any delay. To improve the filtering efficiency, the message list can be adjusted dynamically. The three procedures of data aggregation are all introduced, and the filtering method is designed in detail in this paper. At last, a series of experiments are simulated to prove our scheme’s performance, and the advantages are analyzed in theory. [ABSTRACT FROM AUTHOR]

A wireless sensor network (WSN) model by analyzing the limitations of existing localization algorithm is proposed in this paper. Specifically, neither the anchor nor the ID is required in the proposed model, and the node function is limited. The hardware function of the node in our model is independent of the MAC protocol. Moreover, based on the proposed WSN model, we design an energy efficient anchor-free localization algorithm for no-identity WSN. Instead of multiple measurements, mathematical calculation is adopted to acquire the position information. Subsequently, the effectiveness of the algorithm is confirmed by rigid theory. The simulation results show that the proposed algorithm can improve the performance of localization and energy efficiency in WSN significantly. [ABSTRACT FROM AUTHOR]