Real-time computing without stable states: a new framework for neural computation based on perturbations

A key challenge for neural modeling is to explain how a continuous stream of multimodal input from a rapidly changing environment can be processed by stereotypical recurrent circuits of integrate-and-fire neurons in real time. We propose a new computational model for real-time computing on time-varying input that provides an alternative to paradigms based on Turing machines or attractor neural networks. It does not require a task-dependent construction of neural circuits. Instead, it is based on principles of high-dimensional dynamical systems in combination with statistical learning theory and can be implemented on generic evolved or found recurrent circuitry. It is shown that the inherent transient dynamics of the high-dimensional dynamical system formed by a sufficiently large and heterogeneous neural circuit may serve as universal analog fading memory. Readout neurons can learn to extract in real time from the current state of such recurrent neural circuit information about current and past inputs that may be needed for diverse tasks. Stable internal states are not required for giving a stable output, since transient internal states can be transformed by readout neurons into stable target outputs due to the high dimensionality of the dynamical system. Our approach is based on a rigorous computational model, the liquid state machine, that, unlike Turing machines, does not require sequential transitions between well-defined discrete internal states. It is supported, as the Turing machine is, by rigorous mathematical results that predict universal computational power under idealized conditions, but for the biologically more realistic scenario of real-time processing of time-varying inputs. Our approach provides new perspectives for the interpretation of neural coding, the design of experiments and data analysis in neurophysiology, and the solution of problems in robotics and neurotechnology.     

基于互信息的生物神经网络功能性连接辨识

生物神经网络（ BNN）功能性连接的辨识方法被广泛地应用于使用BNN的多通道时间序列数据构建网络连接结构,帮助加深对BNN结构和功能间关系的认识和理解。首先,建立基于积分点火（ IF）机制的BNN模型,获得多通道神经元脉冲序列;然后,运用互信息（ MI）方法计算出各神经元间的MI值,超过一定阈值的MI表明两个神经元间存在相互连接关系。仿真结果表明：基于MI的网络辨识方法计算开销较小,对BNN功能性连接结构具有较高的辨识度。     

网络表示学习的研究与发展

网络表示学习旨在将网络中的节点表示成低维稠密且具有一定推理能力的向量,以运用于节点分类、社区发现和链路预测等社交网络应用任务中,是连接网络原始数据和网络应用任务的桥梁。传统的网络表示学习方法都是针对网络中节点和连边只有一种类型的同质信息网络的表示学习方法,而现实世界中的网络往往是具有多种节点和连边类型的异质信息网络。而且,从时间维度上来看,网络是不断变化的。因此,网络表示学习的研究方法随着网络数据的复杂化而不断变化。对近年来针对不同网络的网络表示学习方法进行了分类介绍,并阐述了网络表示学习的应用场景。     

Two-Tier genetic programming: towards raw pixel-based image classification

Classifying images is of great importance in machine vision and image analysis applications such as object recognition and face detection. Conventional methods build classifiers based on certain types of image features instead of raw pixels because the dimensionality of raw inputs is often too large. Determining an optimal set of features for a particular task is usually the focus of conventional image classification methods. In this study we propose a Genetic Programming (GP) method by which raw images can be directly fed as the classification inputs. It is named as Two-Tier GP as every classifier evolved by it has two tiers, the other for computing features based on raw pixel input, one for making decisions. Relevant features are expected to be self-constructed by GP along the evolutionary process. This method is compared with feature based image classification by GP and another GP method which also aims to automatically extract image features. Four different classification tasks are used in the comparison, and the results show that the highest accuracies are achieved by Two-Tier GP. Further analysis on the evolved solutions reveals that there are genuine features formulated by the evolved solutions which can classify target images accurately.     

Data stream synchronization for defining meaningful fMRI classification problems

Application of machine learning techniques to the functional Magnetic Resonance Imaging (fMRI) data is recently an active field of research. There is however one area which does not receive due attention in the literature – preparation of the fMRI data for subsequent modelling. In this study we focus on the issue of synchronization of the stream of fMRI snapshots with the mental states of the subject, which is a form of smart filtering of the input data, performed prior to building a predictive model. We demonstrate, investigate and thoroughly discuss the negative effects of lack of alignment between the two streams and propose an original data-driven approach to efficiently address this problem. Our solution involves casting the issue as a constrained optimization problem in combination with an alternative classification accuracy assessment scheme, applicable to both batch and on-line scenarios and able to capture information distributed across a number of input samples lifting the common simplifying i.i.d. assumption. The proposed method is tested using real fMRI data and experimentally compared to the state-of-the-art ensemble models reported in the literature, outperforming them by a wide margin.     

Skewed distributions in semi-stream joins: How much can caching help?

Semi-stream join algorithms join a fast data stream with a disk-based relation. This is important, for example, in real-time data warehousing where a stream of transactions is joined with master data before loading it into a data warehouse. In many important scenarios, the stream input has a skewed distribution, which makes certain performance optimizations possible.We propose two such optimization techniques: (1) a caching technique for frequently used master data and (2) a technique for selective load shedding of stream tuples. The caching technique is fine-grained, operating on a tuple-level. Furthermore, it is generic in the sense that it can be applied to different semi-stream join algorithms to deal with data skew. We analyze it by combining it with various well-known semi-stream joins, and show that it improves the service rate by more than 40% for typical data with skewed distributions. The load shedding technique sheds the fraction of the stream that is most expensive to join. In contrast to existing approaches, the service rate improves under load shedding. We present experimental data showing significant improvements as compared to related approaches and perform a sensitivity analysis for various internal parameters.     

应用Diophantine方程的多通道最优去卷

应用时域上的现代时间序列分析方法,提出了具有ARMA新息滤波器形式的多通 道最优去卷估值器.它要求解一个Diophantine方程,可处理非平稳输入信号,且可统一处理 最优去卷滤波、平滑和预报问题.仿真例子说明了其有效性.     

PGG: An Online Pattern Based Approach for Stream Variation Management

Many database applications require efficient processing of data streams with value variations and fluctuant sampling frequency.The variations typically imply fundamental features of the stream and important domain knowledge of underlying objects.In some data streams,successive events seem to recur in a certain time interval,but the data indeed evolves with tiny differences as time elapses.This feature,so called pseudo periodicity,poses a new challenge to stream variation management.This study focuses on the online management for variations over such streams.The idea can be applied to many scenarios such as patient vital signal monitoring in medical applications.This paper proposes a new method named Pattern Growth Graph (PGG) to detect and manage variations over evolving streams with following features:1) adopts the wave-pattern to capture the major information of data evolution and represent them compactly; 2) detects the variations in a single pass over the stream with the help of wave-pattern matching algorithm;3) only stores different segments of the pattern for incoming stream,and hence substantially compresses the data without losing important information;4) distinguishes meaningful data changes from noise and reconstructs the stream with acceptable accuracy. Extensive experiments on real datasets containing millions of data items,as well as a prototype system,are carried out to demonstrate the feasibility and effectiveness of the proposed scheme.     

基于值分解的多目标多智能体深度强化学习方法

多智能体深度强化学习方法可应用于真实世界中需要多方协作的场景,是强化学习领域内的研究热点。在多目标多智能体合作场景中,各智能体之间具有复杂的合作与竞争并存的混合关系,在这些场景中应用多智能体强化学习方法时,其性能取决于该方法是否能够充分地衡量各智能体之间的关系、区分合作和竞争动作,同时也需要解决高维数据的处理以及算法效率等应用难点。针对多目标多智能体合作场景,在QMIX模型的基础上提出一种基于目标的值分解深度强化学习方法,并使用注意力机制衡量智能体之间的群体影响力,利用智能体的目标信息实现量两阶段的值分解,提升对复杂智能体关系的刻画能力,从而提高强化学习方法在多目标多智能体合作场景中的性能。实验结果表明,相比QMIX模型,该方法在星际争霸2微观操控平台上的得分与其持平,在棋盘游戏中得分平均高出4.9分,在多粒子运动环境merge和cross中得分分别平均高出25分和280.4分,且相较于主流深度强化学习方法也具有更高的得分与更好的性能表现。     

深度学习中对抗样本的构造及防御研究

随着深度学习技术在计算机视觉、网络安全、自然语言处理等领域的进一步发展,深度学习技术逐渐暴露了一定的安全隐患。现有的深度学习算法无法有效描述数据本质特征,导致算法面对恶意输入时可能无法给出正确结果。以当前深度学习面临的安全威胁为出发点,介绍了深度学习中的对抗样本问题,梳理了现有的对抗样本存在性解释,回顾了经典的对抗样本构造方法并对其进行了分类,简述了近年来部分对抗样本在不同场景中的应用实例,对比了若干对抗样本防御技术,最后归纳对抗样本研究领域存在的问题并对这一领域的发展趋势进行了展望。     

Large-Scale Gaussian Process Inference with Generalized Histogram Intersection Kernels for Visual Recognition Tasks

We present new methods for fast Gaussian process (GP) inference in large-scale scenarios including exact multi-class classification with label regression, hyperparameter optimization, and uncertainty prediction. In contrast to previous approaches, we use a full Gaussian process model without sparse approximation techniques. Our methods are based on exploiting generalized histogram intersection kernels and their fast kernel multiplications. We empirically validate the suitability of our techniques in a wide range of scenarios with tens of thousands of examples. Whereas plain GP models are intractable due to both memory consumption and computation time in these settings, our results show that exact inference can indeed be done efficiently. In consequence, we enable every important piece of the Gaussian process framework—learning, inference, hyperparameter optimization, variance estimation, and online learning—to be used in realistic scenarios with more than a handful of data.     

Non-linear observer design by transformation into a generalized observer canonical form

An observer design method for a certain class of non-linear single output systems is introduced. The characteristic feature of this method consists of the fact that it does not require any linearization in the way that the given non-linear system is approximated by a linear one. The present paper deals with the derivation of a transformation of the considered non-linear system into a generalized observer canonical form (GOCF) which enables a systematic observer design similar to the linear one based on the well-known linear observer canonical form. To assign conditions for its existence, the transformation into the GOCF is carried out in two steps via a generalized observability canonical form (GOBCF). In contrast to previous forms used for linear systems, the two non-linear canonical forms presented here also depend on the time derivatives of the input variables. This means that the resulting observer has to be supplied, not only with the input and output variables of the given system, but also with derivatives of the input variables. However, a final example shows that it is possible to eliminate those derivatives in special cases.     

A Learning‐Theoretic Approach To Model‐Set Identification

A new approach to model‐set identification is proposed based on an agnostic learning theory. The squared prediction error is estimated together with its uncertainty uniformly in some parameter region. Based on this estimation, a model set is constructed so as to include the best model. The proposed approach does not require assumptions on the true dynamics or the noise, neither does it need infinite number of input‐output data in order to justify its result. But it guarantees that the size of the identified model set converges to zero as the number of input‐output data increases. Improvement of the precision is considered on the proposed identification method. Generalization of the approach is discussed and a numerical example is presented.     

Efficient Similarity Search over Future Stream Time Series

With the advance of hardware and communication technologies, stream time series is gaining ever-increasing attention due to its importance in many applications such as financial data processing, network monitoring, Web click-stream analysis, sensor data mining, and anomaly detection. For all of these applications, an efficient and effective similarity search over stream data is essential. Because of the unique characteristics of the stream, for example, data are frequently updated and real-time response is required, the previous approaches proposed for searching through archived data may not work in the stream scenarios. Especially, in the cases where data often arrive periodically for various reasons (for example, the communication congestion or batch processing), queries on such incomplete time series or even future time series may result in inaccuracy using traditional approaches. Therefore, in this paper, we propose three approaches, polynomial, Discrete Fourier Transform (DFT), and probabilistic, to predict the unknown values that have not arrived at the system and answer similarity queries based on the predicted data. We also apply efficient indexes, that is, a multidimensional hash index and a B⁺-tree, to facilitate the prediction and similarity search on future time series, respectively. Extensive experiments demonstrate the efficiency and effectiveness of our methods for prediction and answering queries.     

面向股票交易分析场景的流式大数据系统测试框架

分布式集群环境使得数据实时计算更为复杂,流式大数据处理系统的正确性难以保障.现有的大数据基准测试框架可以测试流式大数据处理系统的性能表现,但是普遍存在应用场景设计简单、评价指标不充分等不足.针对这一挑战,本文构造了一个面向股票交易场景的流式大数据基准测试框架,通过生成股票高频交易数据,测试系统在高流速场景下的延迟、吞吐量、GC时间、CPU资源等的性能表现.同时,通过横向测试验证流式大数据系统的扩展性.本文以Apache Spark Streaming为待测系统进行测试,实验结果表明,高流速场景下出现延迟增加、GC时间提高等性能下降问题,原因是系统输入速率的提高及并行度的增加.     

面向数据流的多任务多核在线学习算法

多任务多核学习已逐渐成为在线学习算法研究的热点。对于数据流的处理,现有的在线学习算法在准确性上有一定的欠缺,因此提出一种新的多任务多核在线学习模型用于提高数据流预测的准确性。在保持多任务多核学习的基础上,将其扩展到在线学习中,从而得到一个新的在线学习算法;同时为输入数据保持一定大小的数据窗口,用较小空间换取数据的完整性。实验部分对核函数的选取以及训练样本集的大小进行了较为详细的分析,通过对UCI数据和实际的机场客流量数据进行分析,很好地保障了流数据处理的准确性及实时性,有一定的实际应用价值。     

4K视频流异构多核的多路分屏方法研究

针对视频多屏显示系统无法依据分屏数量进行超高清视频自适应分屏传输的问题,提出了异构多核的视频流传输方法。采用基于ARM的嵌入式系统实现多任务处理与实时监控;通过FPGA实现对视频流进行接收、转换、处理及分屏输出显示的硬件加速。系统可以根据输入信号实时配置FPGA的工作参数,实现分辨率大小、分屏数量可变的超高清视频流分屏显示。测试结果表明,4K视频多路分屏画面拼接无明显错位、同步一致,较好地满足了视频多路分屏显示要求。     

基于多模态的在线序列极限学习机研究

单一模态包含的物体信息有限,导致在物体材质识别分类中表现不佳,而传统多模态融合方法在样本训练过程中需要输入所有数据。提出一种多模态的多尺度局部感受野在线序列极限学习机方法。对物体不同模态样本运用改进的特征提取框架,利用多尺度局部感受野感知样本信息提取特征,并将不同模态特征融合后通过在线序列极限学习机进行训练学习。在线序列极限学习机在训练过程中增量式地输入样本进行训练,当有新数据需要训练时无需对所有数据重新训练。在TUM触觉纹理数据库上进行验证,实验结果表明,多模态融合的分类精度高于单模态的分类精度,且改进的特征提取框架可以显著提升分类性能。