基于Octave卷积的混合精度神经网络量化方法

浮点数位宽的深度神经网络需要大量的运算资源,这导致大型深度神经网络难以在低算力场景（如边缘计算）上部署。为解决这一问题,提出一种即插即用的神经网络量化方法,以压缩大型神经网络的运算成本,并保持模型性能指标不显著下降。首先,基于Octave卷积将输入特征图的高频和低频成分进行分离;其次,分别对高低频分量应用不同位宽的卷积核进行卷积运算;第三,使用不同位宽的激活函数将高低频卷积结果量化至相应位宽;最后,混合不同精度的特征图来获得该层卷积结果。实验结果证实了所提方法压缩模型的有效性,在CIFAR-10/100数据集上,将模型压缩至1+8位宽时,该方法可保持准确率指标的下降小于3个百分点;在ImageNet数据集上,使用该方法将ResNet50模型压缩至1+4位宽时,其正确率指标仍高于70%。     

基于非零和随机博弈的APT攻击主动防御策略选取

为解决APT（高级持续性威胁）攻防对抗过程中的防御滞后性问题,并在有限资源下做出最优主动防御决策,针对APT攻击过程中攻防双方意图、可行策略集随攻击阶段推进而演变的特点进行了研究,基于非合作博弈理论构建了多阶段APT攻防随机博弈模型AO-ADSG（APT-oriented attack-defense stochastic game）。针对APT攻防对抗中双方效用不对等的现象引入非零和思想,设计符合APT攻击特征的全资产要素效用量化方法;在分析博弈均衡的基础上给出最优防御策略选取算法。最后,通过“夜龙攻击”模拟实验验证了提出方法的可行性及正确性。     

Design of forging process variables under uncertainties

Forging is a complex nonlinear process that is vulnerable to various manufacturing anomalies, such as variations in billet geometry, billet/die temperatures, material properties, and workpiece and forging equipment positional errors. A combination of these uncertainties could induce heavy manufacturing losses through premature die failure, final part geometric distortion, and reduced productivity. Identifying, quantifying, and controlling the uncertainties will reduce variability risk in a manufacturing environment, which will minimize the overall production cost. In this article, various uncertainties that affect the forging process are identified, and their cumulative effect on the forging tool life is evaluated. Because the forging process simulation is time-consuming, a response surface model is used to reduce computation time by establishing a relationship between the process performance and the critical process variables. A robust design methodology is developed by incorporating reliability-based optimization techniques to obtain sound forging components. A case study of an automotive-component forging-process design is presented to demonstrate the applicability of the method.     

Connectionist simulation of quantification skills

The study of numerical abilities, and how they are acquired, is being used to explore the continuity between ontogenesis and environmental learning. One technique that proves useful in this exploration is the artificial simulation of numerical abilities with neural networks, using different learning paradigms to explore development. A neural network simulation of subitization, sometimes referred to as visual enumeration, and of counting, a recurrent operation, has been developed using the so-called multi-net architecture. Our numerical ability simulations use two or more neural networks combining supervised and unsupervised learning techniques to model subitization and counting. Subitization has been simulated using networks employing unsupervised self-organizing learning, the results of which agree with infant subitization experiments and are comparable with supervised neural network simulations of subitization reported in the literature. Counting has been simulated using a multi-net system of supervised static and recurrent backpropagation networks that learn their individual tasks within an unsupervised, competitive framework. The developmental profile of the counting simulation shows similarities to that of children learning to count and demonstrates how neural networks can learn how to be combined together in a process modelling development.     

基于样本密度和分类误差率的增量学习矢量量化算法研究

作为一种简单而成熟的分类方法, K最近邻(K nearest neighbor, KNN)算法在数据挖掘、模式识别等领域获得了广泛的应用, 但仍存在计算量大、高空间消耗、运行时间长等问题. 针对这些问题, 本文在增量学习型矢量量化(Incremental learning vector quantization, ILVQ)的单层竞争学习基础上, 融合样本密度和分类误差率的邻域思想, 提出了一种新的增量学习型矢量量化方法, 通过竞争学习策略对代表点邻域实现自适应增删、合并、分裂等操作, 快速获取原始数据集的原型集, 进而在保障分类精度基础上, 达到对大规模数据的高压缩效应. 此外, 对传统近邻分类算法进行了改进, 将原型近邻集的样本密度和分类误差率纳入到近邻判决准则中. 所提出算法通过单遍扫描学习训练集可快速生成有效的代表原型集, 具有较好的通用性. 实验结果表明, 该方法同其他算法相比较, 不仅可以保持甚至提高分类的准确性和压缩比, 且具有快速分类的优势.     

工业控制系统风险评估要素量化方法

风险评估是保障工业控制系统安全的重要手段之一,在风险分析原理的基础上,对其中的资产、安全措施、威胁、脆弱性等关键要素给出了相应的量化方法。通过对相关风险要素进行细致的量化,有利于降低风险评估过程中的主观性,提高评估结果的合理性、准确性。     

Error quantification of the normalised right graph symbol for an errors-in-variables system

This paper proposes a novel method to quantify the error of a nominal normalized right graph symbol (NRGS) for an errors-in-variables (EIV) system corrupted with bounded noise. Following an identification framework for estimation of a perturbation model set, a worst-case v-gap error bound for the estimated nominal NRGS can be first determined from \textit{a priori} and \textit{a posteriori} information on the underlying EIV system. Then, an NRGS perturbation model set can be derived from a close relation between the v-gap metric of two models and ${\rm H}_\infty$-norm of their NRGSs' difference. The obtained NRGS perturbation model set paves the way for robust controller design using an ${\rm H}_\infty$ loop-shaping method because it is a standard form of the well-known NCF (normalized coprime factor) perturbation model set. Finally, a numerical simulation is used to demonstrate the effectiveness of the proposed identification method.     

Suboptimal robust synthesis for MIMO plant under coprime factor perturbations

Suboptimal robust synthesis for MIMO nominal system under coprime factor perturbations is considered in classical and non-classical statements. In the classical statement, weights of perturbations and upper bound on magnitude bounded exogenous disturbance are assumed to be known to controller designer. Suboptimal synthesis within ε tolerance is reduced to the solution of log₂(1/ε) standard mixed sensitivity problems of ℓ₁ optimization. In the non-classical statement, the upper bounds on perturbations and exogenous disturbance are to be estimated from measurement data and suboptimal synthesis is reduced to the solution of 1/ε mixed sensitivity problems.     

Characterization of surface deformation with the Edge of LightTM technique

The aircraft lap joints are inspected with an enhanced visual inspection technique named “Edge of Light”, which is patented by the NRC Institute for Aerospace Research. This technique is applicable for rapid detection of possible hidden corrosion in lap joints. The surface deformation due to hidden corrosion can be characterized by this optical-based inspection method. In this study, a calibration procedure is developed to quantify the lap joint surface deformation. The effect of surface reflectivity is investigated with the solid film highlighting technique (SolidHi^TM), which helps achieve a uniform reflectivity during the inspection. The efficiency of the technique is demonstrated with the experimental results. This work is supported by NRC-IAR New Initiative Research Funding.