Currently, many problems such as variable signal resources, complex execution environments, and low efficiency of scheduling algorithms are faced by heterogeneous signal processing platforms. The task scheduling algorithm is one of the key factors that directly affect the performance of the processing platform. In order to solve the problems of low efficiency of task scheduling algorithms and high computational cost of processors, a heterogeneous platform scheduling algorithm based on the quantum genetic algorithm is proposed in this paper. The algorithm constructs a task scheduling model by using a directed acyclic graph. This paper quantifies the mapping relationship between the quantum genetic algorithm and task scheduling. It corresponds qubits to binary, chromosomes to processor numbers, and individuals to processor scheduling strategies. In this paper, a new way of coding chromosomes using quantum coherence properties is designed to reduce the population size and increase population diversity. Crossover operations are performed on all individuals using full-interference crossover to avoid the results falling into local optimal solutions. The population of slow convergence is solved by implementing mutation operations on populations through quantum rotation gates. In addition, a task pre-ordering stage is designed based on the table scheduling algorithm. The task scheduling priority developed at this stage is used as the reference value for the initial encoding of the population, so that the search space for solutions is reduced. Finally, experiments are conducted using randomly generated task graphs. The algorithm is compared with improved genetic algorithms and existing intelligent scheduling algorithms. The results show that the algorithm can still obtain better results when the number of populations and iterations is small. It is more appropriate for heterogeneous platforms and computation-intensive tasks.
Semi-supervised graph learning aims to infer the class of unlabeled nodes or graphs by using various prior knowledge in a given graph.By improving the automation of data labeling,semi-supervised graph learning has high efficiency in node classification,but as a deep learning architecture,it also faces the threat of backdoor attacks,but no effective backdoor attack method has been developed for semi-supervised graph node classification tasks.This study propose a persistent clean-label backdoor attack method for semi-supervised graph node classification models,which generates poisoned samples by adaptively adding triggers and perturbations on unlabeled training data,and then trains to obtain poisoned semi-supervised graph node classification models without modifying the labels.The attacker can poison the model more stealthily with a poisoning rate no higher than 4％.To ensure the persistence of the backdoor in the model,a hyperparameter tuning strategy is also proposed to select the optimal value of the per-turbation.Extensive experiments on several semi-supervised graph node classification models and open-source datasets show that the proposed approach achieves an attack success rate of up to 96.25％with little loss of classifica-tion accuracy of the model on normal samples.
Most medical image segmentation methods mainly focus on training and evaluating in the same or similar medical data domain,which need lots of pixel-level annotations.However,these models face challenges in out-of-distribution medical data set,which is known as"domain shift"problem.A fixed U-shaped segmentation structure is usually used to solve this problem,resulting in it not being better adapted to specific partition tasks.A gradient-based meta-learning and neural architecture search method is proposed in this study,which can adjust the segmentation network according to specific tasks to achieve good performance and have good generalization ability.This method mainly uses the specific task to carry out the architecture search module to further improve the segmenta-tion effect,and then uses the gradient-based meta-learning training algorithm to improve the generalization ability.On the public dataset M&Ms,under the 5%label data,its Dice and Hausdorff distance are 79.62%and 15.38%.Under 2%label data,its Dice and Hausdorff distance are 74.03%and 17.05%.Compared with other mainstream methods,the proposed method has better generalization ability.
传统故障诊断方法存在编码不完整和编码边界过于绝对等缺点,难以满足电网运维工作的实际需求.利用电力变压器故障时产生的气体对变压器进行故障诊断是目前智能电网状态检测的热门研究领域.但变压器出现各类故障的频率具有较大差别,可能造成故障样本信息不完整、数据量不足,使用传统卷积神经网络模型会出现训练过程收敛速度慢、训练精度低等问题.文中提出一种在保留原始数据特征不变的情况下进行数据增强的方法,将扩充后的一维数据转化为二维图片输入到二维卷积神经网络诊断模型中,并对卷积神经网络模型中的Adam优化算法进行改进.诊断结果表明,所提方法使得网络训练的精度达到了96.20%,相较于传统的一维卷积神经网络故障诊断方法(92.12%)具有更高的收敛速度和泛化能力.
在集成剪枝中,为了同时高效地选择优质、独立的分类器,文中提出了一种新的动态自适应交叉策略的遗传算法用于分类器的集成剪枝.该方法使用轮盘赌和贪婪策略动态更新每个交叉策略的优先级,根据优先级计算每个策略被选中的概率,从而在算法迭代过程中自适应选择不同的交叉策略.此外,该方法考虑了交叉概率和变异概率动态自适应变化,并使用大多数投票法对挑选出来的分类器进行集成以获得最终结果.将文中所提方法与一些集成剪枝方法在5 个真实人脸表情数据集上进行对比,实验结果表明文中所提该方法可以选出效果更好、冗余度更低的分类器,并在CK +数据集上具有22.50%的最低误差.
