Recently, stereo matching models based on 3D CNNs have achieved excellent performance. However, huge computational burden and memory footprint of deep 3D convolution limit their deployment on edge devices and real-time scenes. Furthermore, given that disparity maps exhibit continuous disparity variations in content regions but sharp disparity variants at edge areas, spatially shared weight mechanism inherent in convolution may struggle to handle both types of regions. In this paper, we propose a lightweight stereo matching model, called DBCANet, with Dual-Branch Cost Aggregation based on region segmentation to reduce the computational burden. Specifically, we divide the cost aggregation module into two branches: the edge branch is to aggregate edge disparity information, and the content branch is to aggregate the disparity information of the content region. 2D convolution reduces the computational burden and dual branch can mitigate the problem caused by convolutional spatially-shared weights to compensate for the accuracy loss. We extensively validate our model on Scene Flow, KITTI datasets, and our model can achieve the lowest multiply-accumulate operations (MACs) among speed-oriented stereo matching models, while its EPE is only 0.59 on Scene Flow dataset, which is ranked second. Our model can achieve a good balance between accuracy and model computation.

At present, signal transmission in mobile communication is susceptible to multipath effects and interference, causing a decrease in spectral efficiency and an increase in energy consumption. However, existing methods cannot effectively synergistically optimize beamforming and power allocation. To address this issue, this paper constructs an intelligent metasurface-assisted communication system signal transmission optimization model based on I-DQNN. This model uses dual intelligent metasurface-assisted base stations as the communication architecture core, and combines GCNN and QNN to achieve channel estimation and beamforming optimization. This study introduces Squeeze-and-Excitation Networks, gradient constraints, and parameter sharing mechanisms to optimize and improve the model. In practical applications, the error rate, communication energy efficiency, average latency, and system throughput of the research method were 0.11%, 4.45 bit/Hz/J, 3.05 ms, and 1005.95 Mbps. During its operation, the system failure rate remained below 0.17%, and the number of communication interruptions was only 17 times. The proposed model can effectively enhance the stability and energy efficiency of signal transmission in complex channel environments, significantly reduce bit error rates and delays, and achieve efficient utilization of spectrum resources in high-density user scenarios.

Weighted Adversarial Training trains models by assigning training weights to distinct data points and computing weighted losses. Existing methods adhere to the principle of ”assigning higher weights to more vulnerable data points” in weight allocation strategies, overlooking the variation in model weight perturbations required for training different data points to their intended targets. Data points demanding substantial model weight adjustments can negatively impact model performance. This paper proposes the logit margin change rate as an evaluation metric for model robustness at individual data points and constructs a weight allocation function via a greedy weighting strategy. By analyzing the degree of weight adjustments needed for data points to achieve training objectives, this strategy prioritizes assigning higher weights to points requiring less perturbation demand. This enables the model to learn more data points within limited expressive capacity, thereby enhancing robustness. Experimental results across multiple datasets demonstrate that our method further improves model robustness beyond existing approaches.

The adversarial attack is characterized by a high attack success rate and a fast generation of examples. It is widely used in neural network robustness evaluation and adversarial training. Restricted by the randomness of the initialization of the attack point and the iterative finding algorithm can not guarantee that it can reach the global optimal solution, the existing adversarial attack methods have attack uncertainty in a single attack and need to increase the number of attacks in order to improve the attack success rate. This paper defines the label susceptibility to analyze the attack effect. For adversarial data with high label susceptibility, using interval analysis to find the adversarial examples in its neighbourhood can effectively alleviate the attack uncertainty and improve the attack success rate. The experimental results demonstrate that our method achieves comparable or superior attack success rates with fewer attack iterations compared to other approaches requiring more iterations. When applied to smaller datasets and with reduced attack scopes, our method exhibits even higher attack efficiency.

In adversarial training, it is important to utilize the limited expressive ability of the model to enhance the learning of vulnerable data points. Conventional adversarial training treats all adversarial data equally in the loss function, which may result in unreasonable allocation of the model’s expressive ability, limiting further robustness enhancement. We propose a dynamic weighting loss method based on robust distance to adjust the decision boundary of the model at data points. It enables the model to prioritize the robustness improvement of the more vulnerable data points in adversarial training. The method quantifies the robustness of the model over data points by robust distances, defining relative distances as adjusted expectations of the model’s decision boundaries at the data points and calculating the adversarial learning priorities for the data points. Experimental results on multiple datasets show that the method can further improve the robustness of the model.

With a massive influx of terminal devices connecting to wireless network, the rapid growth in wireless network traffic poses significant challenges to existing network architectures. Device-to-Device (D2D) communication facilitates direct data exchanges between devices, reducing latency and network congestion. However, implementing D2D communication technology in edge caching poses privacy and security concerns. This paper proposes a privacy-preserving edge caching algorithm based on permissioned blockchain and federated reinforcement learning (PBFRL) to ensure data privacy and security. The PBFRL algorithm incorporates a permissioned blockchain deployed on both base stations and user devices to ensure data privacy and security. Considering the varied hardware resources of user devices, an asynchronous training method is introduced, separating the sample collection algorithm and the updating network algorithm to run on different devices. Simulation results demonstrate that the proposed algorithm surpasses other benchmark algorithms in terms of content transmission delay and data privacy and security.

Contrastive learning has been widely applied in domain adaptation tasks due to its powerful representation learning capabilities in recent years. Constructing positive and negative sample pairs through different augmented views of instances promotes the effective application of contrastive learning in domain adaptation. However, most existing methods are based on Noise Contrastive Estimation (NCE), which primarily focuses on instance-level information and lacks attention to the relationships between different instances. These methods treat all negative samples as noise interference, leading to an increased distance between negative samples and anchor samples, disregarding instances in the negative samples that share the same category as the anchor sample. Consequently, these methods do not offer significant advantages in domain adaptation for image classification tasks. To address this issue, we propose a method called Consistent Positive Correlation Sample Distribution (CPCSD) to mitigate the problem of class collision among negative samples by leveraging the semantic similarity between instances. We introduce a positive correlation distribution loss within the contrastive adaptation framework to compute the similarity distribution between batches of augmented views and align them. Additionally, we sharpen the target similarity distribution to obtain amore emphasized relationship distribution, thereby alleviating the issue of negative sample noise in contrastive adaptation. Extensive experiments demonstrate that our proposed model has significant advantages in contrastive adaptation algorithms and improves the performance of downstream domain adaptation tasks.

Blockchain technology has garnered significant attention from academia and industry, with scalability remaining a key challenge. Sharding is a promising solution, dividing the blockchain into smaller partitions called shards, each processing a portion of the transactions to increase throughput. This approach is critical for enabling efficient Proof of Stake (PoS) consensus mechanisms, as demonstrated by the transition of Dogecoin to PoS, where sharding reduces the computational burden on validators and enhances scalability. However, sharding introduces high storage redundancy, as nodes in each shard must collectively maintain a copy of the entire blockchain, imposing substantial storage pressure. To address this, segments are introduced to divide the main chain into smaller parts distributed across nodes. Existing methods, however, randomly assign segments to nodes, resulting in high costs for node setup and segment queries. This paper investigates the optimal allocation of segments within shards to minimize these costs, proposing a Segment Allocation algorithm based on Cost Clustering (SACC). Theoretical analysis and simulations demonstrate that SACC achieves lower setup, query, and total costs while maintaining security and scalability, offering a more efficient solution for sharding-based PoS blockchains like Dogecoin.

Edge caching can effectively reduce service latency, alleviate backhaul traffic pressure, and enhance the quality of experience (QoE) through the reuse of popular network content. Many studies utilize deep reinforcement learning (DRL) methods to learn optimal caching strategies for mobile edge networks. However, traditional DRL methods often suffer from limitations such as inefficient sample utilization, lengthy training periods, and insufficient model generalization performance, which necessitate relearning network parameters when faced with new tasks. To overcome these challenges, this paper proposes a meta-reinforcement learning based fast adaptive edge caching algorithm (MRFAC). It models the mobile edge caching problem as a markov decision process (MDP) and designs a DRL algorithm to optimize the cache replacement strategy, aiming to maximize long-term caching benefits. Moreover, meta-learning is incorporated into DRL algorithm, treating the learning process of DRL as an independent learning objective. The weights of the gated recurrent unit (GRU) are used to encode and store the learned prior knowledge, enabling MRFAC to rapidly adapt to new tasks. Experiment results indicate that the cache performance of MRFAC outperforms baseline methods such as LFU, FIFO, and DDPG. Furthermore, MRFAC exhibits rapid adaptation to new tasks, with significantly improved convergence speed compared to traditional DRL methods.

Single-pixel imaging is an advanced computational imaging technique based on compressive sensing that offers higher signal-to-noise ratio and broader application scope compared to traditional imaging techniques. However, conventional reconstruction algorithms suffer from issues such as long processing time and low reconstruction accuracy during the sampling and reconstruction processes. Deep learning-based compressed reconstruction networks can circumvent the complex iterative computations of traditional algorithms and achieve fast, high-quality reconstruction. In this paper, we propose a Variational Multi-Scale Error Compensation Network (VMSE) based on variational autoencoders. VMSE designs an error compensation network to enhance the feature representation capability of the sampling reconstruction network. We employ multiple latent variables to generate error features at different scales in the intermediate layers of the error compensation network, compensating the reconstructed image. Additionally, we design a module that simultaneously learns in the spatial and frequency domains, which is used for upsampling and complementing the missing high-frequency information in the frequency domain. On the MNIST dataset, when the sampling rate is 0.025, VMSE achieved higher Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity Index(SSIM) scores, especially with an SSIM score of 0.963, significantly surpassing Reconnet and DR2Net's scores of 0.930 and 0.920, respectively. This was further corroborated by practical experiments, where at low sampling rates, VMSE could reconstruct outlines more clearly compared to TVAL3.

Sharding presents an enticing path toward improving blockchain scalability. However, the consensus mechanism within individual shards faces mounting security challenges due to the restricted number of consensus nodes and the reliance on conventional, unchanging nodes for consensus. Common strategies to enhance shard consensus security often involve increasing the number of consensus nodes per shard. While effective in bolstering security, this approach also leads to a notable rise in consensus delay within each shard, potentially offsetting the scalability advantages of sharding. Hence, it becomes imperative to strategically select nodes to form dedicated consensus groups for each shard. These groups should not only enhance shard consensus security but also do so without exacerbating consensus delay. In this article, we propose a novel consensus group selection based on transmission delay between nodes (CGSTD) to address this challenge, with the goal of minimizing the overall consensus delay across the system. CGSTD intelligently selects nodes from various shards to form distinct consensus groups for each shard, thereby enhancing shard security while maintaining optimal system-wide consensus efficiency. We conduct a rigorous theoretical analysis to evaluate the security properties of CGSTD and derive approximation ratios under various operational scenarios. Simulation results validate the superior performance of CGSTD compared to baseline algorithms, showcasing reductions in total consensus delay, mitigated increases in shard-specific delay, optimized block storage utilization per node, and streamlined participation of nodes in consensus groups.

In recent years, many deep models for stereo matching employ 3D convolutions for cost aggregation to achieve better performance. However, this approach requires substantial computational and memory resources, limiting the deployment of the model on edge devices. In this paper, we analyze the challenges of channel compression in stereo matching models and adopt a straightforward and efficient compression method. Our method focuses on the channel compression of the cost aggregation module, enabling the model to achieve acceleration on existing hardware, reducing significantly computational cost. We set a hyperparameter κ that decides the compression rate, and compress the channels of each layer in cost aggregation module according to it. This method ensures the consistency for irregular skip connections. We extensively test our method on GwcNet, PSMNet and CFNet on multiple datasets, achieving promising results. We believe that the proposed compression method allows stereo matching models to better balance computational cost and accuracy during various degrees of channel compression, making them more suitable for deployment on edge devices.

The increasing size of pre-trained language models has led to a growing interest in model compression. Pruning and distillation are the primary methods employed to compress these models. Existing pruning and distillation methods are effective in maintaining model accuracy and reducing its size. However, they come with limitations. For instance, pruning is often suboptimal and biased by transforming it into a continuous optimization problem. Distillation relies primarily on one-to-one layer mappings for knowledge transfer, which leads to underutilization of the rich knowledge in teacher. Therefore, we propose a method of joint pruning and distillation for automatic pruning of pre-trained language models. Specifically, we first propose Gradient Progressive Pruning (GPP), which achieves a smooth transition of indicator vector values from real to binary by progressively converging the values of unimportant units’ indicator vectors to zero before the end of the search phase. This effectively overcomes the limitations of traditional pruning methods while supporting compression with higher sparsity. In addition, we propose the Dual Feature Distillation (DFD). DFD adaptively globally fuses teacher features and locally fuses student features, and then uses the dual features of global teacher features and local student features for knowledge distillation. This realizes a “preview-review” mechanism that can better extract useful information from multi-level teacher information and transfer it to student. Comparative experiments on the GLUE benchmark dataset and ablation experiments indicate that our method outperforms other state-of-the-art methods.

Domain adaptive object detection (DAOD) is an effective approach to solving the domain shift problem, which aims to improve the generalization ability on the target domain through joint training with labeled source domain and unlabeled target domain data. However, in the real world, the source domain is often inaccessible due to privacy, legal, and regulatory reasons. Source-free object detection (SFOD) adapts a well-trained source domain object detector to the target domain without needing access to source domain data. Most existing SFOD methods are based on a self-training framework (such as the teacher–student framework), where the teacher model guides the student model’s training by setting a confidence threshold to filter out unreliable pseudo-labels. However, this may lead to the loss of semantic information in the target domain, resulting in suboptimal model performance. Therefore, in this paper, we propose a contrastive learning framework guided by proposal-level reliable feature guidance. Specifically, we first build a general self-training framework based on Faster-RCNN. Then, we introduce a novel proposal-level pseudo-label annotation and filtering method (PAF), improving the annotation and filtering of pseudo-labels and obtaining proposal-level reliable features. Next, we design reliable feature banks to categorize and store reliable proposal features. Finally, we propose classification contrastive learning (CCL) and bounding box contrastive learning (BCL) to further extract semantic information from reliable proposal features, learning the lost semantic information in the target domain. Extensive experiments on seven benchmark datasets demonstrate the effectiveness of our method.

Peak Age of Information(PAoI), as a performance indicator representing the freshness of information, has attracted the attention of researchers in recent years. The data packet transmission rate in the LoRa network determines the information freshness level for system packets. In order to study the optimal scheduling of data packets, we try to use the PAoI to describe the real-time level of the end devices( $$ EDs $$ ) and reduce it. We use edge servers to process monitoring data packets at the edge of the network to improve the efficiency of $$ EDs $$ and the information freshness level of data. Since packet transmission will be constrained by $$ EDs $$ battery queue energy and gateway queue backlog, we propose an optimization problem that aims to minimize the long-term average PAoI of $$ EDs $$ while ensuring network stability. With the Lyapunov optimization framework, the long-term stochastic optimization problem is transformed into a single-slot optimization problem. Furthermore, to avoid the problem of too large search space, we propose a dynamic strategy space reduction algorithm (SSDR) to shrink the strategy space. The simulation experiments show that our SSDR algorithm can optimize the PAoI index of $$ EDs $$ in various situations and satisfy the constraints of long-term optimization.

Sharding is a promising strategy to enhance blockchain scalability. However, the surge in transactions has led to heightened relations between nodes in the system, reflecting the volume of transactions between them. The increase in related nodes engaging in identical transactions across diverse shards leads to substantial cross-shard transactions, contributing to communication delays and impeding enhancements in throughput. Current methods typically employ greedy or heuristic approaches to organize nodes into shards, resulting in marginal reductions in the total relation between related nodes in different shards (i.e., the number of cross-shard transactions), while causing shard imbalance. Hence, there is a crucial need for periodic shard reorganization based on node relations to minimize the total relation between related nodes across different shards while ensuring shard balance. In this article, we investigate the reorganization of nodes into shards based on node relations in sharding blockchains, aiming to minimize the total relation between related nodes in different shards.We formulate the shard reorganization problem and introduce the shard reorganization algorithm based on the relation between nodes (SRRN) to address this issue. Theoretical analysis proves that SRRN is a 2 lambda M-approximation algorithm, where lambda=(r(max)/r(min)), with M representing the number of shards, and r(max) and r(min) denoting the maximum and minimum nonzero relations between nodes, respectively. Simulation results demonstrate that SRRN outperforms baseline algorithms in terms of total relation, degree of relation reduction, differences in computing power between shards, cross-shard ratio, and throughput.

Image clustering is a crucial task in remote sensing data processing. Nevertheless, relying solely on traditional machine learning algorithms is insufficient for processing large amounts of images and accurately clustering them. Some recent works have successfully applied deep learning to remote sensing image clustering, significantly improving clustering performance. However, current deep clustering methods struggle to correctly assign boundary samples that are close to multiple clusters, as their features are very similar to samples in different clusters. To this end, we propose a method called Pseudo Dual-view Contrastive Learning via Taylor Expansion Optimization for Remote Sensing Image Clustering, namely, PDTE. Specifically, we introduce the pseudo dual-view contrastive learning framework with random aggressive augmentation to explore the potential information and subtle feature differences of images, and optimize its objective function using Taylor expansion to enlarge the difference in distance between boundary samples and each cluster. This allows us to assign boundary samples accurately and improves the clustering performance. Extensive experiments conducted on five datasets show that PDTE has superior clustering performance.

Large-scale pre-trained models, such as BERT, have demonstrated outstanding performance in Natural Language Processing (NLP). Nevertheless, the high number of parameters in these models has increased the demand for hardware storage and computational resources while posing a challenge for their practical deployment. In this article, we propose a combined method of model pruning and knowledge distillation to compress and accelerate large-scale pre-trained language models. Specifically, we introduce a dynamic structure pruning method based on differentiable search and recursive knowledge distillation to automatically prune the BERT model, named DDK. We define the search space for network pruning as all feed-forward layer channels and self-attention heads at each layer of the network, and utilize differentiable methods to determine their optimal number. Additionally, we design a recursive knowledge distillation method that employs adaptive weighting to extract the most important features from multiple intermediate layers of the teacher model and fuse them to supervise the student network learning. Our experimental results on the GLUE benchmark dataset and ablation analysis demonstrate that our proposed method outperforms other advanced methods in terms of average performance.

Weight and activation binarization can efficiently compress deep neural networks and accelerate model inference, but they cause severe accuracy degradation. Existing optimization methods for binary neural networks (BNNs) focus on fitting full-precision networks to reduce quantization errors and suffer from a tradeoff between accuracy and efficiency. In contrast, considering information loss and the mismatch between model capacity and input information quantity caused by network binarization, we propose Information Restriction and Information Recovery Network (IR $$^2$$ Net) to stimulate the potential of BNNs and achieve improved network accuracy by restricting the input information and recovering feature information. The proposed approach includes (1) information restriction, which evaluates the feature information extracted from the input by a BNN, discards some of the information it cannot focus on, and reduces the amount of the input information to match the model capacity; and (2) information recovery: due to the information loss incurred during forward propagation, the extracted feature information of the network is not sufficient for supporting accurate classification. Shallow feature maps with richer information are selected, and these feature maps are fused with the final feature maps to recover the extracted feature information and further enhance the model capacity to match the amount of input information. In addition, the computational cost is reduced by streamlining the information recovery method to strike a better tradeoff between accuracy and efficiency. Experimental results demonstrate that our approach still achieves comparable accuracy even with a $$\sim$$ 10x floating-point operations (FLOPs) reduction for ResNet-18. The models and code are available at https://github.com/pingxue-hfut/IR2Net .

Network binarization (i.e., binary neural networks, BNNs) can efficiently compress deep neural networks and accelerate model inference but cause severe accuracy degradation. Existing BNNs are mainly implemented based on the commonly used full-precision network backbones, and then the accuracy is improved with various techniques. However, there is a question of whether the full-precision network backbone is well adapted to BNNs. We start from the factors of the performance degradation of BNNs and analyze the problems of directly using full-precision network backbones for BNNs: for a given computational budget, the backbone of a BNN may need to be shallower and wider compared to the backbone of a full-precision network. With this in mind, Depth-Width Reshaping (DWR) is proposed to reshape the depth and width of existing full-precision network backbones and further optimize them by incorporating pruning techniques to better fit the BNNs. Extensive experiments demonstrate the analytical result and the effectiveness of the proposed method. Compared with the original backbones, the DWR backbones constructed by the proposed method result in close to O(√s) decrease in activations, while achieving an absolute accuracy increase by up to 1.7% with comparable computational cost. Besides, by using the DWR backbones, existing methods can achieve new state-of-the-art (SOTA) accuracy (e.g., 67.2% on ImageNet with ResNet-18 as the original backbone). We hope this work provides a novel insight into the backbone design of BNNs. The code is available at https://github.com/pingxue-hfut/DWR.