Low-Rank Adaptation (LoRA) enables parameter-efficient fine-tuning of Large Language Models (LLMs), and recent Mixture-of-Experts (MoE) extensions further enhance flexibility by dynamically combining multiple LoRA experts.However, existing MoE-augmented LoRA methods assume that experts operate independently, often leading to unstable routing, expert dominance.In this paper, we propose TalkLoRA, a communication-aware MoELoRA framework that relaxes this independence assumption by introducing expert-level communication prior to routing.TalkLoRA equips low-rank experts with a lightweight Talking Module that enables controlled information exchange across expert subspaces, producing a more robust global signal for routing.Theoretically, we show that expert communication smooths routing dynamics by mitigating perturbation amplification while strictly generalizing existing MoELoRA architectures.Empirically, TalkLoRA consistently outperforms vanilla LoRA and MoELoRA across diverse language understanding and generation tasks, achieving higher parameter efficiency and more balanced expert routing under comparable parameter budgets.These results highlight structured expert communication as a principled and effective enhancement for MoE-based parameterefficient adaptation.Code is available at https://github.com/why0129/TalkLoRA.

Frequency Fitness Assignment (FFA) offers an alternative take on metaheuristic optimization. Here, the encounter frequencies of objective values are used to make the selection decisions. This leads to a variety of interesting algorithm features, such as an invariance under all injective transformations of the objective function value and a very strong focus on exploration of the search space. In this article, for the first time, we discover a condition under which purely FFA-guided search can actually get stuck, even on a problem as simple as OneMax. To tackle this issue, we suggest hybrid approaches combining objective-guided and FFA-guided search. We propose using crossover for solution transfer between the two component algorithms of the hybrids. Our experiments show that (1) the original FFA allows us to solve problems like Trap, TwoMax, and Jump in (experimentally observed) polynomial time; (2) the suggested hybrids address FFA's shortcomings and are occasionally orders of magnitude faster; and (3) we report several new best-known solutions for the NP-hard low-autocorrelation binary sequences problem.

Accurate prediction of lithium-ion battery SOH is critical for the safety and lifespan extension of a battery system. Feature noise and insufficient model accuracy remain major obstacles. To address these challenges, this study proposes a correlation-driven outlier detection and reconstruction method grounded in multidimensional feature engineering to enhance SOH prediction. First, from five dimensions—constant current charging, constant voltage charging, incremental capacity, temperature, and unit time voltage change rate—this work systematically defines 45 degradation features, with 33 applicable to CALCE and 38 to NASA datasets due to dataset-specific signal availability. Second, for each target feature, it screens highly correlated feature subsets using the Pearson correlation coefficient. A linear mapping is then established through regression analysis to model the expected values, enabling identification and reconstruction of outliers. Finally, based on the reconstructed feature dataset, this study conducts comparative experiments to systematically evaluate the effects of different noise-filtering strategies on model prediction accuracy and robustness. Experimental results showed that the proposed method significantly improved both the accuracy and robustness of SOH prediction. On the CALCE dataset, relative to raw data, the average RMSE, MAE, and MAPE decreased by 9.40%, 4.30%, and 6.30%, respectively. Compared with the traditional 3σ filtering method, the improvements reached 10.60%, 3.78%, and 5.63%, respectively. On the NASA dataset, the standalone gain was smaller; however, when combined with 3σ filtering, the average improvements in RMSE, MAE, and MAPE reached about 8.5%, exceeding the 3σ filtering alone by approximately 7%. Comprehensive SHAP analysis further revealed that, compared with the 3σ filtering scheme, the features processed by CorODR exhibited more stable and interpretable importance distributions.

Low-Light Image Enhancement (LLIE) methods based on either Retinex theory or deep learning still exhibit significant shortcomings in handling image corruptions, such as noise, artifacts, and color distortion. The primary issue is that both Retinex algorithms and existing networks may introduce or amplify these corruptions during enhancement. To address these limitations, we propose the Denoised-Modulated Hybrid-Semantic Scale-Aware Network (DHSNet), a novel one-stage LLIE method. DHSNet integrates a Signal-to-Noise Ratio (SNR)-based denoising mechanism and a Hybrid-Semantic Scale-Aware Module (HSM) to preprocess noise and fuse multi-scale features for robust image enhancement. Moreover, we introduce the Illumination Partial Attention Block (IPAB) to further improve illumination correction and nonlinear transformation capabilities. DHSNet effectively mitigates noise, preserves intricate details, and restores degraded structures. Extensive experiments on multiple LLIE datasets demonstrate that it outperforms state-of-the-art (SOTA) methods in both qualitative and quantitative metrics. Furthermore, DHSNet exhibits strong generalization in no-reference LLIE and low-light object detection tasks, underscoring its practical value for real-world applications.

Low-rank adaptation (LoRA) has been developed as an efficient approach for adapting large language models (LLMs) by finetuning two low-rank matrices, thereby reducing the number of trainable parameters.However, prior research indicates that many of the weights in these matrices are redundant, leading to inefficiencies in parameter utilization.To address this limitation, we introduce Dense Low-Rank Adaptation (DenseLoRA), a novel approach that enhances parameter efficiency while achieving superior performance compared to LoRA.DenseLoRA builds upon the concept of representation fine-tuning, incorporating a single Encoder-Decoder to refine and compress hidden representations across all adaptation layers before applying adaptation.Instead of relying on two redundant low-rank matrices as in LoRA, DenseLoRA adapts LLMs through a dense low-rank matrix, improving parameter utilization and adaptation efficiency.We evaluate DenseLoRA on various benchmarks, showing that it achieves 83.8% accuracy with only 0.01% of trainable parameters, compared to LoRA's 80.8% accuracy with 0.70% of trainable parameters on LLaMA3-8B.Additionally, we conduct extensive experiments to systematically assess the impact of DenseLoRA's components on overall model performance.Code is available at https://github.com/mulin-ahu/DenseLoRA.

Deep neural networks (DNNs) often exhibit out-of-distribution (OOD) overconfidence, producing overly confident predictions on OOD samples. We attribute this issue to the inherent over-complexity of DNNs and investigate two key aspects: capacity and nonlinearity. First, we demonstrate that reducing model capacity through knowledge distillation can effectively mitigate OOD overconfidence. Second, we show that selectively reducing nonlinearity by removing ReLU operations further alleviates the issue. Building on these findings, we present a practical guide to model simplification, combining both strategies to significantly reduce OOD overconfidence. Extensive experiments validate the effectiveness of this approach in mitigating OOD overconfidence and demonstrate its superiority over state-of-the-art methods. Additionally, our simplification strategies can be combined with existing OOD detection techniques to further enhance OOD detection performance.

Adversarial attacks on point clouds are crucial for assessing and improving the adversarial robustness of 3D deep learning models. Traditional solutions strictly limit point displacement during attacks, making it challenging to balance imperceptibility with adversarial effectiveness. In this paper, we attribute the inadequate imperceptibility of adversarial attacks on point clouds to deviations from the underlying surface. To address this, we introduce a novel point-to-surface (P2S) field that adjusts adversarial perturbation directions by dragging points back to their original underlying surface. Specifically, we use a denoising network to learn the gradient field of the logarithmic density function encoding the shape’s surface, and apply a distance-aware adjustment to perturbation directions during attacks, thereby enhancing imperceptibility. Extensive experiments show that adversarial attacks guided by our P2S field are more imperceptible, outperforming state-of-the-art methods.

ABSTRACT Real‐time monitoring of fires is crucial for safeguarding lives and property. However, current fire detection methods still suffer from issues such as redundant feature information, poor network generalisation capabilities and low perception of target location information. To address these challenges, a novel fire detection method called YOLO‐FDI has been proposed. This method utilises partial convolution and coordinate convolution with attention mechanisms and Alpha loss at different stages. Specifically, to enhance target localisation accuracy, an attention mechanism is integrated into the model to autonomously focus on fire‐affected areas. In terms of feature extraction, partial convolution is employed to reduce computational redundancy and memory access, improving performance and effectively extracting spatial features. During the feature fusion stage, coordinate convolution embeds feature information into coordinate data, further enhancing the coordinate perception capabilities of pixels on the feature map, thereby improving adaptability and accuracy in detecting fire targets. Additionally, the model utilises Alpha loss to enhance flexibility and robustness in fire object detection and recognition. Experimental results demonstrate the effectiveness of the proposed model based on three self‐constructed datasets. Compared to the baseline YOLOv7 model, its mAP has improved by 4.5 percentage points, 1.7 percentage points and 2.6 percentage points, respectively. This method demonstrates the capability to accurately represent fire targets and exhibits better stability and reliability in fire target detection, effectively reducing false positives and missed detections.

To address the substantial coordination overhead and communication latency inherent in cross-shard transaction commits within contemporary multi-shard blockchain architectures, this paper presents an efficient cross-shard atomic commit scheme (PledgeACS), grounded in the use of pledge transactions. The proposed scheme introduces a pledge transaction mechanism that employs a null recipient address and synchronizes these transactions across shards to an auxiliary chain through a global consensus protocol. Additionally, a cross-shard transaction protocol is developed, securing recipient funds via pledge transactions within the global consensus framework. Furthermore, a batch pledge transaction record and settlement protocol tailored for shard blockchains is designed, followed by rigorous feasibility analysis and performance evaluation. Experimental results indicate that the proposed scheme markedly decreases the user-perceived latency in cross-shard transactions and enhances security relative to current solutions, providing an innovative approach for achieving high throughput and scalability in blockchain systems.

The Quadratic Assignment Problem (QAP) is one of the classical N P-hard tasks from operations research with a history of more than 65 years. It is often approached with heuristic algorithms and over the years, a multitude of such methods has been applied. All of them have in common that they tend to prefer better solutions over worse ones. We approach the QAP with Frequency Fitness Assignment (FFA), an algorithm module that can be plugged into arbitrary iterative heuristics and that removes this bias. One would expect that a heuristic that does not care whether a new solution is better or worse compared to the current one should not perform very well. We plug FFA into a simple randomized local search (RLS) and yield the FRLS, which surprisingly outperforms RLS on the vast majority of the instances of the well-known QAPLIB benchmark set.

The Traveling Salesperson Problem (TSP) is one of the most well-known N P-hard optimization tasks. A randomized local search (RLS) is not a good approach for solving TSPs, as it quickly gets stuck at local optima. FRLS, the same algorithm with Frequency Fitness Assignment plugged in, has been shown to be able to solve many more TSP instances to optimality. However, it was also assumed that its performance will decline if an instance has a large number M of different possible objective values. How can we explore these more or less obvious algorithm properties in a controlled fashion, if determining the number #L of local optima or the size BL of their joint basins of attraction as well as the feature M are N P-hard problems themselves? By creating TSP instances with a small number of cities for which we can actually know these features! We develop a deterministic construction method for creating TSP instances with rising numbers M and a sampling based approach for the other features. We determine all the instance features exactly and can clearly confirm the obvious (in the case of RLS) or previously suspected (in the case of FRLS) properties of the algorithms. Furthermore, we show that even with small-scale instances, we can make interesting new findings, such as that local optima seemingly have little impact on the performance of FRLS.

The classical compact double-round robin traveling tournament problem (TTP) asks us to schedule the games of n teams in a tournament such that each team plays against every other team twice, once at home and once away (doubleRoundRobin constraint). The maxStreak constraint prevents teams from having more than three consecutive home or away games. The noRepeat constraint demands that, before two teams can play against each other the second time, they must at least play one other game in between. The goal is to find a game plan observing all of these constraints and having the overall shortest travel length. We define a gamepermutation based encoding that allows for representing game plans with arbitrary numbers of constraint violations and tackle the TTP as a bi-objective problem minimizing both the number of constraint violations and the travel length by applying the well-known NSGA-II. We combine both objectives in a lexicographic prioritization scheme and also apply the randomized local search RLS to this single-objective variant of the problem. We realize that Frequency Fitness Assignment (FFA), which makes algorithms invariant under all injective transformations of the objective function value, would also make optimization algorithms invariant under all lexicographic prioritization schemes for multi-objective problems. The FRLS, i.e., the RLS with FFA plugged in, would therefore solve both possible prioritizations of our TTP variants at once. We thus also explore its performance on the TTP. We find that RLS performs surprisingly well and can find game plans without constraint violations reliably until a scale of 36 teams, whereas FRLS and NSGA-II have an advantage on small- and mid-scale problems.

We consider a two-dimensional orthogonal bin packing problem (2BP) where rectangular items are to be placed into rectangular bins such that their edges are parallel to those of the bins with the aim to require as few bins as possible. Two variants of the problem are analyzed. In the 2BP|O|F, the items have a fixed orientation while in the 2BP|R|F, they can be rotated by 90 degrees. We show that on both variants, a simple randomized local search (RLS) has surprisingly good performance – if the objective function guiding the search is defined suitably. In particular, on the 2BP|O|F, the RLS performs on par with more complicated state-of-the-art metaheuristics. We furthermore investigate plugging Frequency Fitness Assignment (FFA) into the RLS, obtaining the FRLS. FFA has improved the RLS performance on several classical N P-hard optimization problems from operations research, including Max-SAT, the Job Shop Scheduling Problem, and the Traveling Salesperson Problem. This paper is the first negative result for FFA: it cannot improve algorithm performance on the 2BP variants studied. This can be explained by the fact that RLS already performs very well on the instances of the 2DPackLib benchmark set used as the basis of our experiments.

Graph Convolutional Networks (GCNs) are widely used for skeleton-based action recognition and achieved remarkable performance. Due to the locality of graph convolution, GCNs can only utilize short-range node dependencies but fail to model long-range node relationships. In addition, existing graph convolution based methods normally use a uniform skeleton topology for all frames, which limits the ability of feature learning. To address these issues, we present the Graph Convolution Network with Self-Attention (SelfGCN), which consists of a mixing features across self-attention and graph convolution (MFSG) module and a temporal-specific spatial self-attention (TSSA) module. The MFSG module models local and global relationships between joints by executing graph convolution and self-attention branches in parallel. Its bi-directional interactive learning strategy utilizes complementary clues in the channel dimensions and the spatial dimensions across both of these branches. The TSSA module uses self-attention to learn the spatial relationships between joints of each frame in a skeleton sequence. It also models the unique spatial features of the single frames. We conduct extensive experiments on three popular benchmark datasets, NTU RGB+D, NTU RGB+D120, and Northwestern-UCLA. The results of the experiment demonstrate that our method achieves or exceeds the record accuracies on all three benchmarks. Our project website is available at https://github.com/SunPengP/SelfGCN.

Abstract Image segmentation plays a pivotal role in image processing. Level set model is a traditional variation image segmentation method. In order to achieve level set evolution equation, the level set energy functionals are minimized with the gradient descent methods and then the partial differential equation (PDE) was solved by finite difference scheme. Slow speed is one of its disadvantages. We propose a sweep optimization algorithm based on global cosine fitting (GCF) energy. Instead of calculating the PDE and the curvature, the proposed sweeping algorithm directly calculates the energy change when a pixel moves from one side of evolving contour to the other. It checks whether the GCF energy is decreased or not. The proposed algorithm has many advantages. For example, independent of initial level set contour positions and parameters, need not consider the Courant Friedrichs Lew condition and the regularization energy term. The proposed algorithm can be easily extended to high dimension image segmentation. The experiments on synthetic images, noise images and real images demonstrate the effectiveness of the proposed sweeping optimization algorithm.

A fitness assignment process transforms the features (such as the objective value) of a candidate solution to a scalar fitness, which then is the basis for selection. Under frequency fitness assignment (FFA), the fitness corresponding to an objective value is its encounter frequency in selection steps and is subject to minimization. FFA creates algorithms that are not biased toward better solutions and are invariant under all injective transformations of the objective function value. We investigate the impact of FFA on the performance of two theory inspired, state-of-the-art evolutionary algorithms, the Greedy (2+1) GA and the self-adjusting$(1+(\lambda,\lambda))$GA. FFA improves their performance significantly on some problems that are hard for them. In our experiments, one FFA-based algorithm exhibited mean runtimes that appear to be polynomial on the theory-based benchmark problems in our study, including traps, jumps, and plateaus. We propose two hybrid approaches that use both direct and FFA-based optimization and find that they perform well. All FFA-based algorithms also perform better on satisfiability problems than any of the pure algorithm variants.

There is a high danger of accidents in chemical manufacturing plants. A devise that could automatically detect safety signs in the vicinity of a person could issue verbal warnings in order to reduce this risk. The most important task here is to correctly identify such signs from images. While there have been many achievements in the field of traffic sign detection, there is only very little research on detecting safety signs. In this work, we first provide an open and comprehensive benchmark dataset with 4650 images (expanded to 27900 images) of 30 chemical safety signs, each with a class label and bounding box and annotated with other image features. We then conduct a comprehensive analysis comparing the performance of the state-of-the-art deep learning models Faster R-CNN, SSD, YOLOv3-spp, and YOLOv5 on this dataset. In our study, YOLOv5 performs the best. It has the best mean average precision (98.9%), the best average recall (96.9%), and can process the highest number of images per second (71) on our hardware. It is already close to be sufficient for real-world application, but, like all investigated methods, suffers when detecting many signs at once, small signs, or signs in front of complex backgrounds. Our study closes an important gap in research and lays the foundation for solid future work in the domain of sign detection for improving worker safety.

Under frequency fitness assignment (FFA), the fitness corresponding to an objective value is its encounter frequency in fitness assignment steps and is subject to minimization. FFA renders optimization processes invariant under bijective transformations of the objective function value. On TwoMax, Jump, and Trap functions of dimension s, the classical (1 + 1)-EA with standard mutation at rate 1/s can have expected runtimes exponential in s. In our experiments, a (1 + 1)-FEA, the same algorithm but using FFA, exhibits mean runtimes that seem to scale as s2ln s. Since Jump and Trap are bijective transformations of OneMax, it behaves identical on all three. On OneMax, LeadingOnes, and Plateau problems, it seems to be slower than the (1 + 1)-EA by a factor linear in s. The (1 + 1)-FEA performs much better than the (1 + 1)-EA on W-Model and MaxSat instances. We further verify the bijection invariance by applying the Md5 checksum computation as transformation to some of the above problems and yield the same behaviors. Finally, we show that FFA can improve the performance of a memetic algorithm for job shop scheduling.