The integration of intelligent reflective surfaces (IRS) on unmanned aerial vehicles (UAVs), termed UAV-IRS, to bolster wireless communications has emerged as a hotspot of academic research and industrial application. In this paper, we investigate the problem of secure communication in the harsh communication environment assisted by multiple UAV-IRSs, where the UAV-IRSs act as relays to assist the downlink secure communication between the base station and the users. To maximize the security sum rate between the base station and the users, the trajectory planning and phase shift design of multiple UAV-IRS needs to be jointly optimized. To solve this complex non-convex optimization problem, we introduce a distributed collaborative optimization scheme for multiple UAV-IRSs called credit-aware cooperative multi-agent reinforcement learning (MARL), which takes MARL as the base algorithm, and then solves the credit allocation problem among multiple UAV-IRSs by using cooperative game theory to facilitate exploration, and finally constrains non-cooperative behaviors among UAV-IRSs by using the primal-dual optimization algorithm to promote cooperation. Finally, the effectiveness and superiority of the proposed scheme is verified by comprehensive simulation experiments.

Large Language Models (LLMs), despite their remarkable capabilities, are hampered by hallucinations.A particularly challenging variant, knowledge overshadowing, occurs when one piece of activated knowledge inadvertently masks another relevant piece, leading to erroneous outputs even with high-quality training data.Current understanding of overshadowing is largely confined to inference-time observations, lacking deep insights into its origins and internal mechanisms during model training.Therefore, we introduce PHANTOMCIRCUIT, a novel framework designed to comprehensively analyze and detect knowledge overshadowing.By innovatively employing knowledge circuit analysis, PHANTOMCIRCUIT dissects the function of key components in the circuit and how the attention pattern dynamics contribute to the overshadowing phenomenon and its evolution throughout the training process.Extensive experiments demonstrate PHANTOMCIRCUIT 's effectiveness in identifying such instances, offering novel insights into this elusive hallucination and providing the research community with a new methodological lens for its potential mitigation.Our code can be found in https://github.com/halfmorepiece/PhantomCircuit.

Satellite networks play a vital role in providing global connectivity to remote areas, including mountains, forests, and regions affected by natural disasters. The primary challenge lies in the limited communication timeframe between satellites and earth stations (ESs) due to the swift motion of satellites, making timely satellite data downloads through ESs challenging. To address this, we propose a method named priority-aware and throughput-optimized intersatellite cooperative data transmission (PACT). PACT optimizes network throughput while maximizing download priorities for ESs by leveraging intersatellite links (ISLs), considering diverse download priorities for different data types. To comprehensively capture constellation characteristics, PACT models low Earth orbit (LEO) constellations using a spatiotemporal graph. Within the graph, data priorities are assigned as edge weights, and the allocation of initial download windows to ESs is achieved by maximum weighted matching. Subsequently, PACT organizes a contact plan through cooperative scheduling leveraging ISLs. A bipartite graph is constructed based on data awaiting download and link allocation to redistribute remaining download windows optimally through maximum matching. This iterative process enhances network throughput and maintains data priority. Performance assessments in the ndnSIM framework, covering diverse load scenarios, demonstrate the efficiency and benefits of PACT, particularly in prioritizing data downloads.

Abstract Catastrophic forgetting in neural networks during incremental learning remains a challenging problem. Previous research investigated catastrophic forgetting in fully connected networks, with some earlier work exploring activation functions and learning algorithms. Applications of neural networks have been extended to include similarity learning. Understanding how similarity learning loss functions would be affected by catastrophic forgetting is of significant interest. Our research investigates catastrophic forgetting for four well-known similarity-based loss functions during incremental class learning. The loss functions are Angular, Contrastive, Center, and Triplet loss. Our results show that the catastrophic forgetting rate differs across loss functions on multiple datasets. The Angular loss was least affected, followed by Contrastive, Triplet loss, and Center loss with good mining techniques. We implemented three existing incremental learning techniques, iCaRL, EWC, and EBLL. We further proposed a novel technique using Variational Autoencoders (VAEs) to generate representation as exemplars passed through the network’s intermediate layers. Our method outperformed three existing state-of-the-art techniques. We show that one does not require stored images (exemplars) for incremental learning with similarity learning. The generated representations from VAEs help preserve regions of the embedding space used by prior knowledge so that new knowledge does not “overwrite” it.

During the check node (CN) update, the elements of input message vectors are redundant for the output message vectors. Hence, in this paper, we exactly select from the input message vectors the elements, which really have contributions to the error-correction performance and constitute the minimum set for the CN update. With adoption of the forward and backward (FB) scheme, an adaptive minimum-set min-sum algorithm (AMSA) is proposed to reduce the computation complexity of the FB process. In order to concurrently update the output vectors belonging to the same CN, we present a parallel minimum-set min-sum algorithm (PMSA) with lower memory complexity than the AMSA. Compared with the min-sum algorithms, the two proposed minimum-set based algorithms introduce no error performance loss.

Intelligent Transportation Systems (ITS) urgently require real-time data collection with minimized Age of Information (AoI), yet face critical challenges from high-dynamic traffic environments and unstable wireless channels. By virtue of the low deployment cost and the high-speed mobility, Uncrewed Aerial Vehicle (UAV) bring us a feasible approach to the aforementioned problem. Nevertheless, such a problem is far from trivial due to lot of factors ranging from the highly dynamic communication environment, the dimension-varying input/output for each UAV, to the extremely large solution space for all the UAVs as a whole in a distributed collaborative manner. Although existing Multi-Agent Reinforcement Learning (MARL) solutions are widely used to address the above challenges, they all rely on fixed-dimensional input/output processing (e.g., padding/truncation strategies), leading to redundancy or loss of information due to dimensionality-changing scenarios. To address this gap, we proposed an improvement scheme based on attention-driven MARL, which redesigns the policy and critic network based on the attention mechanism to help UAVs extract critical information from dimension-varying data in an accurate and efficient manner. Finally, we verify the superiority and robustness of the proposed scheme through multiple sets of experiments with multiple different aspects. The simulation results show that the proposed scheme is scalable and efficient, and the weighted average AoI under different scenarios is lower than the existing state-of-the-art schemes by$13.1\%\sim 56.2\%$.

Projecting visual features into word embedding space has become a significant fusion strategy adopted by Multimodal Large Language Models (MLLMs).However, its internal mechanisms have yet to be explored.Inspired by multilingual research, we identify domain-specific neurons in multimodal large language models.Specifically, we investigate the distribution of domain-specific neurons and the mechanism of how MLLMs process features from diverse domains.Furthermore, we propose a threestage mechanism for language model modules in MLLMs when handling projected image features, and verify this hypothesis using logit lens.Extensive experiments indicate that while current MLLMs exhibit Visual Question Answering (VQA) capability, they may not fully utilize domain-specific information.Manipulating domain-specific neurons properly will result in a 10% change of accuracy at most, shedding light on the development of cross-domain, all-encompassing MLLMs in the future.The source code is available at this URL.

With the blossom of Internet of Things (IoT) technology, the big data volumes raised by the large number of IoT devices have posed great burden on the communication and computing network. Considering the advantages of reflecting intelligent surface (RIS), mobile edge computing (MEC), and unmanned aerial vehicle (UAV), this article proposes a flying-RIS-assisted MEC system to assist offloading services to alleviate the ground computation burden in IoT. The UAV equipped with RIS is dispatched to fly over a specific area to assist in offloading the ground’s computing mission to MEC server situated nearby access point (AP) in IoT. The cost of the IoT device is introduced as the weighted sum of the device’s energy consumption and the time consumed to accomplish all computation tasks. To prolong the lifetime and guarantee the communication quality of the IoT devices, this article minimizes the sum cost of all IoT devices by collaboratively planning UAV’s trajectory, scheduling the IoT devices’ association with flying-RIS, and optimizing the phase shift value of each reflecting components. To address the posed nonconvex optimization challenge, a deep deterministic policy gradient (DDPG)-based algorithm is brought forward. Besides, the state and action normalization mechanism is used to ease up on the training difficulty. Finally, the numerical simulation results prove the superiority of the proposed algorithm compared with other benchmark schemes.

Stylistic analysis enables open-ended and exploratory observation of languages. To fill the gap in the quantitative analysis of the stylistic systems of Middle Chinese, we construct lexical features based on the evolutive core word usage and scheme a Bayesian method for feature parameters estimation. The lexical features are from the Swadesh list, each of which has different word forms along with the language evolution during the Middle Ages. We thus count the varied word of those entries along with the language evolution as the linguistic features. With the Bayesian formulation, the feature parameters are estimated to construct a high-dimensional random feature vector to obtain the pair-wise dissimilarity matrix of all the texts based on different distance measures. Finally, we perform the spectral embedding and clustering to visualize, categorize, and analyze the linguistic styles of Middle Chinese texts. The quantitative result agrees with the existing qualitative conclusions and, furthermore, betters our understanding of the linguistic styles of Middle Chinese from both the inter-category and intra-category aspects. It also helps unveil the special styles induced by the indirect language contact.

Mobile edge computing (MEC) has attracted widespread attention due to its ability to effectively alleviate the cloud computing load and significantly reduce latency. However, the potential eavesdroppers challenge the security of the MEC systems and the rapid development of artificial intelligence (AI) has made this security situation more severe. In most existing studies, the eavesdroppers are non-intelligent and it is assumed that they are fixed or move in a simple manner. Obviously, there is a gap from such an assumption to the real conditions that the eavesdropping unmanned aerial vehicles (UAVs) may adjust their flight paths intelligently. To better reflect real-world scenarios, we consider a multi-UAV-assisted MEC system in the presence of intelligent eavesdroppers and propose an adversarial multi-agent reinforcement learning (MARL)-based scheme for secure computational offloading and resource allocation. With this scheme, we aim to solve the zero-sum game between the legitimate UAVs and the eavesdropping UAVs, in which the two types of UAVs take turns acting as the agents of MARL to alternately optimize their respective opposing objectives. The simulation experimental results indicate that the proposed scheme significantly outperforms the existing baseline methods in dealing with the intelligent eavesdropping UAVs, and ensures high energy efficiency of Internet of Things (IoT) devices even in the worst-case scenario when dealing with potential eavesdropping threats.

Computation offloading tends to be an effective way for mitigating computing pressure of user equipments (UEs). By computation offloading, the task can be handled in network edge and/or cloud center to compensate insufficient resources and capabilities of UEs. In this study, we construct a three-tier cloud-edge integration network, where user tasks are offloaded to satellite based edge server and further to the remote ground cloud server via backhaul links. The optimization problem is modeled for minimizing system energy consumption and considers user association, power allocation, task scheduling, and bandwidth assignment jointly. By the proposed schemes based on relaxation transformation and fractional programming, four subproblems are transformed into corresponding convex optimization problems and solved respectively. In order to find the global optimal solutions, a joint iterative algorithm for three-tier computation offloading problem is designed. In numerical simulations, we compare different communication schemes and computation offloading schemes to present the rationality and superiority of the designed algorithm for reducing system energy consumption.

Mobile-edge computing (MEC) as a promising technology enables it to satisfy ever-increasing demands for low-latency and ultrareliable services. However, due to the limitations of computing capability and the dynamic network environment, it is challenging to process massive data with low latency. In this article, we consider a dynamic MEC network with a high-performance edge server, multiple time-varying channels, and multiple mobile devices. We aim to find a policy that can maximize the processing success rate of computational tasks and the fairness index of the system while minimizing the process delays. To this end, we propose a heuristic-assisted multiagent reinforcement learning (RL)-based framework to realize the joint optimization of computation offloading and resource allocation. On the one hand, heuristic search is introduced in this framework to find a better resource allocation policy in edge servers and further assist the multiagent RL algorithm to determine offloading policy in mobile devices. On the other hand, a novel parameterized multiagent RL algorithm based on soft actor–critic (SAC) is also proposed to broaden the effectiveness and availability of the proposed framework. Simulation results of the average cumulative reward, success rate, processing delay, and fairness index fully verify the superiority of the proposed framework and algorithm for supporting this problem.

This paper tackles face recognition in videos employing metric learning methods and similarity ranking models. The paper compares the use of the Siamese network with contrastive loss and Triplet Network with triplet loss implementing the following architectures: Google/Inception architecture, 3D Convolutional Network (C3D), and a 2-D Long short-term memory (LSTM) Recurrent Neural Network. We make use of still images and sequences from videos for training the networks and compare the performances implementing the above architectures. The dataset used was the YouTube Face Database designed for investigating the problem of face recognition in videos. The contribution of this paper is two-fold: to begin, the experiments have established 3-D Convolutional networks and 2-D LSTMs with the contrastive loss on image sequences do not outperform Google/Inception architecture with contrastive loss in top n rank face retrievals with still images. However, the 3-D Convolution networks and 2-D LSTM with triplet Loss outperform the Google/Inception with triplet loss in top n rank face retrievals on the dataset; second, a Support Vector Machine (SVM) was used in conjunction with the CNNs' learned feature representations for facial identification. The results show that feature representation learned with triplet loss is significantly better for n-shot facial identification compared to contrastive loss. The most useful feature representations for facial identification are from the 2-D LSTM with triplet loss. The experiments show that learning spatio-temporal features from video sequences is beneficial for facial recognition in videos.