Neural networks and perceptual learning

Sensory perception is a learned trait. The brain strategies we use to perceive the world are constantly modified by experience. With practice, we subconsciously become better at identifying familiar objects or distinguishing fine details in our environment. Current theoretical models simulate some properties of perceptual learning, but neglect the underlying cortical circuits. Future neural network models must incorporate the top-down alteration of cortical function by expectation or perceptual tasks. These newly found dynamic processes are challenging earlier views of static and feedforward processing of sensory information.     

Neurons compute internal models of the physical laws of motion

A critical step in self-motion perception and spatial awareness is the integration of motion cues from multiple sensory organs that individually do not provide an accurate representation of the physical world. One of the best-studied sensory ambiguities is found in visual processing, and arises because of the inherent uncertainty in detecting the motion direction of an untextured contour moving within a small aperture. A similar sensory ambiguity arises in identifying the actual motion associated with linear accelerations sensed by the otolith organs in the inner ear. These internal linear accelerometers respond identically during translational motion (for example, running forward) and gravitational accelerations experienced as we reorient the head relative to gravity (that is, head tilt). Using new stimulus combinations, we identify here cerebellar and brainstem motion-sensitive neurons that compute a solution to the inertial motion detection problem. We show that the firing rates of these populations of neurons reflect the computations necessary to construct an internal model representation of the physical equations of motion.     

Bottlenose dolphins perceive object features through echolocation

How organisms (including people) recognize distant objects is a fundamental question. The correspondence between object characteristics (distal stimuli), like visual shape, and sensory characteristics (proximal stimuli), like retinal projection, is ambiguous. The view that sensory systems are 'designed' to 'pick up' ecologically useful information is vague about how such mechanisms might work. In echolocating dolphins, which are studied as models for object recognition sonar systems, the correspondence between echo characteristics and object characteristics is less clear. Many cognitive scientists assume that object characteristics are extracted from proximal stimuli, but evidence for this remains ambiguous. For example, a dolphin may store 'sound templates' in its brain and identify whole objects by listening for a particular sound. Alternatively, a dolphin's brain may contain algorithms, derived through natural endowments or experience or both, which allow it to identify object characteristics based on sounds. The standard method used to address this question in many species is indirect and has led to equivocal results with dolphins. Here we outline an appropriate method and test it to show that dolphins extract object characteristics directly from echoes.     

Non-classical receptive field mediates switch in a sensory neuron's frequency tuning

Animals have developed stereotyped communication calls to which specific sensory neurons are well tuned. These communication calls must be discriminated from environmental signals such as those produced by prey. Sensory systems might have evolved neural circuitry to encode both categories. In weakly electric fish, prey and communication signals differ in their spatial extent and frequency content. Here we show that stimuli of different spatial extents mimicking prey and communication signals cause a switch in the frequency tuning and spike-timing precision of electrosensory pyramidal neurons, resulting in the selective and optimal encoding of both stimulus categories. As in other sensory systems, pyramidal neurons respond only to stimuli located within a restricted region of space known as the classical receptive field (CRF). In some systems, stimulation outside the CRF but within a non-classical receptive field (nCRF) can modulate the neural response to CRF stimulation even though nCRF stimulation alone fails to elicit responses. We show that pyramidal neurons possess a nCRF and that it can modulate the response to CRF stimuli to induce this neurobiological switch in frequency tuning.     

基于三线摆法和Origin7．0软件测定重力加速度新方法的研究

本实验针对三线摆的特征，设计出重力加速度测定的新方法，利用Origin7．0软件的线性拟合来处理实验数据，从而能更准确地计算出舟山＇-3地的重力加速度。     

Fluid particle accelerations in fully developed turbulence

The motion of fluid particles as they are pushed along erratic trajectories by fluctuating pressure gradients is fundamental to transport and mixing in turbulence. It is essential in cloud formation and atmospheric transport, processes in stirred chemical reactors and combustion systems, and in the industrial production of nanoparticles. The concept of particle trajectories has been used successfully to describe mixing and transport in turbulence, but issues of fundamental importance remain unresolved. One such issue is the Heisenberg-Yaglom prediction of fluid particle accelerations, based on the 1941 scaling theory of Kolmogorov. Here we report acceleration measurements using a detector adapted from high-energy physics to track particles in a laboratory water flow at Reynolds numbers up to 63,000. We find that, within experimental errors, Kolmogorov scaling of the acceleration variance is attained at high Reynolds numbers. Our data indicate that the acceleration is an extremely intermittent variable--particles are observed with accelerations of up to 1,500 times the acceleration of gravity (equivalent to 40 times the root mean square acceleration). We find that the acceleration data reflect the anisotropy of the large-scale flow at all Reynolds numbers studied.     

The temporal response of the brain after eating revealed by functional MRI

After eating, the human brain senses a biochemical change and then signals satiation, but precisely when this occurs is unknown. Even for well-established physiological systems like glucose-insulin regulation, the timing of interaction between hormonal processes and neural events is inferred mostly from blood sampling. Recently, neuroimaging studies have provided in vivo information about the neuroanatomical correlates of the regulation of energy intake. Temporal orchestration of such systems, however, is crucial to the integration of neuronal and hormonal signals that control eating behaviour. The challenge of this functional magnetic resonance imaging study is to map not only where but also when the brain will respond after food ingestion. Here we use a temporal clustering analysis technique to demonstrate that eating-related neural activity peaks at two different times with distinct localization. Importantly, the differentiated responses are interacting with an internal signal, the plasma insulin. These results support the concept of temporal parcellation of brain activity, which reflects the different natures of stimuli and responses. Moreover, this study provides a neuroimaging basis for detecting dynamic processes without prior knowledge of their timing, such as the acute effects of medication and nutrition in the brain.     

A sensory source for motor variation

Suppose that the variability in our movements is caused not by noise in the motor system itself, nor by fluctuations in our intentions or plans, but rather by errors in our sensory estimates of the external parameters that define the appropriate action. For tasks in which precision is at a premium, performance would be optimal if no noise were added in movement planning and execution: motor output would be as accurate as possible given the quality of sensory inputs. Here we use visually guided smooth-pursuit eye movements in primates as a testing ground for this notion of optimality. In response to repeated presentations of identical target motions, nearly 92% of the variance in eye trajectory can be accounted for as a consequence of errors in sensory estimates of the speed, direction and timing of target motion, plus a small background noise that is observed both during eye movements and during fixations. The magnitudes of the inferred sensory errors agree with the observed thresholds for sensory discrimination by perceptual systems, suggesting that the very different neural processes of perception and action are limited by the same sources of noise.     

非线性信道判决反馈递归神经网络自适应均衡器

针对存在严重符号间串扰及轻度非线性畸变的数字信道，提出了一种判决反馈递归神经网络为均衡器，它将传统的线性信道判决反馈结构巧妙地融入递归神经中，在自适应训练时用期待２信号代替判决反馈信号。     

Nonlinear Analyses of the Dynamic Properties of Hydrostatic Bearing Systems

Nonlinear analyses of hydrostatic bearing systems are necessary to adequately model the fluid-solid interaction. The dynamic properties of linear and nonlinear analytical models of hydrostatic bearings are compared in this paper. The analyses were based on the determination of the aperiodic border of transient processes with external step loads. The results show that the dynamic properties can be most effectivelyimproved by increasing the hydrostatic bearing crosspiece width and additional pocket volume in a bearing can extend the load range for which the transient process is aperiodic, but an additional restrictor and capacitor (RC) chain must be introduced for increasing damping. The nonlinear analyses can also be used to predict typical design parameters for a hydrostatic bearing.     

Neural network computation with DNA strand displacement cascades

The impressive capabilities of the mammalian brain--ranging from perception, pattern recognition and memory formation to decision making and motor activity control--have inspired their re-creation in a wide range of artificial intelligence systems for applications such as face recognition, anomaly detection, medical diagnosis and robotic vehicle control. Yet before neuron-based brains evolved, complex biomolecular circuits provided individual cells with the 'intelligent' behaviour required for survival. However, the study of how molecules can 'think' has not produced an equal variety of computational models and applications of artificial chemical systems. Although biomolecular systems have been hypothesized to carry out neural-network-like computations in vivo and the synthesis of artificial chemical analogues has been proposed theoretically, experimental work has so far fallen short of fully implementing even a single neuron. Here, building on the richness of DNA computing and strand displacement circuitry, we show how molecular systems can exhibit autonomous brain-like behaviours. Using a simple DNA gate architecture that allows experimental scale-up of multilayer digital circuits, we systematically transform arbitrary linear threshold circuits (an artificial neural network model) into DNA strand displacement cascades that function as small neural networks. Our approach even allows us to implement a Hopfield associative memory with four fully connected artificial neurons that, after training in silico, remembers four single-stranded DNA patterns and recalls the most similar one when presented with an incomplete pattern. Our results suggest that DNA strand displacement cascades could be used to endow autonomous chemical systems with the capability of recognizing patterns of molecular events, making decisions and responding to the environment.     

Human cerebellar activity reflecting an acquired internal model of a new tool

Theories of motor control postulate that the brain uses internal models of the body to control movements accurately. Internal models are neural representations of how, for instance, the arm would respond to a neural command, given its current position and velocity. Previous studies have shown that the cerebellar cortex can acquire internal models through motor learning. Because the human cerebellum is involved in higher cognitive function as well as in motor control, we propose a coherent computational theory in which the phylogenetically newer part of the cerebellum similarly acquires internal models of objects in the external world. While human subjects learned to use a new tool (a computer mouse with a novel rotational transformation), cerebellar activity was measured by functional magnetic resonance imaging. As predicted by our theory, two types of activity were observed. One was spread over wide areas of the cerebellum and was precisely proportional to the error signal that guides the acquisition of internal models during learning. The other was confined to the area near the posterior superior fissure and remained even after learning, when the error levels had been equalized, thus probably reflecting an acquired internal model of the new tool.     

The perception of moving plaids reveals two motion-processing stages

When viewed through a small aperture, the perceived motion exhibited by a long moving line or grating is ambiguous. This situation prevails because even a perfect machine could only detect motion perpendicular to a moving contour, so motion parallel to a contour is undetectable. The human visual system views the world through an aperture array--the neural receptive fields. Therefore a moving object is viewed through many small apertures and the motion within many of those apertures is ambiguous. This ambiguity may be resolved by monitoring the motion of a distinctive feature, such as a line-end or corner, and attributing to the larger object the motion of the feature. Alternatively, Adelson and Movshon have suggested that moving images are processed in two stages, that is, they are first decomposed into one-dimensional components which are later recombined to generate perceived object motion. For a moving plaid, defined as the sum of two drifting gratings, these alternative models generate different predictions concerning the resolution of the plaid's motion ambiguity. A feature monitor would respond to the motion of the intersections between gratings, whereas the two-stage motion processor would first decompose the plaid into its constituent gratings and subsequently recombine them to generate the perception of a moving plaid. Using speed discrimination to distinguish between the two models, I find that discrimination thresholds reflect the speed of a plaid's component gratings, rather than the speed of the plaid itself. This result supports the two-stage model. Although speed discrimination is limited by component processing, observers cannot directly access component speed. The only perceptually accessible velocity signal is generated by the second-stage pattern processing.     

Unconscious priming eliminates automatic binding of colour and alphanumeric form in synaesthesia

Synaesthesia is an unusual perceptual phenomenon in which events in one sensory modality induce vivid sensations in another. Individuals may 'taste' shapes, 'hear' colours, or 'feel' sounds. Synaesthesia was first described over a century ago, but little is known about its underlying causes or its effects on cognition. Most reports have been anecdotal or have focused on isolated unusual cases. Here we report an investigation of 15 individuals with colour-graphemic synaesthesia, each of whom experiences idiosyncratic but highly consistent colours for letters and digits. Using a colour-form interference paradigm, we show that induced synaesthetic experiences cannot be consciously suppressed even when detrimental to task performance. In contrast, if letters and digits are presented briefly and masked, so that they are processed but unavailable for overt report, the synaesthesia is eliminated. These results show that synaesthetic experiences can be prevented despite substantial processing of the sensory stimuli that otherwise trigger them. We conclude that automatic binding of colour and alphanumeric form in synaesthesia arises after initial processes of letter and digit recognition are complete.     

Perception of sniff phase in mouse olfaction

Olfactory systems encode odours by which neurons respond and by when they respond. In mammals, every sniff evokes a precise, odour-specific sequence of activity across olfactory neurons. Likewise, in a variety of neural systems, ranging from sensory periphery to cognitive centres, neuronal activity is timed relative to sampling behaviour and/or internally generated oscillations. As in these neural systems, relative timing of activity may represent information in the olfactory system. However, there is no evidence that mammalian olfactory systems read such cues. To test whether mice perceive the timing of olfactory activation relative to the sniff cycle ('sniff phase'), we used optogenetics in gene-targeted mice to generate spatially constant, temporally controllable olfactory input. Here we show that mice can behaviourally report the sniff phase of optogenetically driven activation of olfactory sensory neurons. Furthermore, mice can discriminate between light-evoked inputs that are shifted in the sniff cycle by as little as 10 milliseconds, which is similar to the temporal precision of olfactory bulb odour responses. Electrophysiological recordings in the olfactory bulb of awake mice show that individual cells encode the timing of photoactivation in relation to the sniff in both the timing and the amplitude of their responses. Our work provides evidence that the mammalian olfactory system can read temporal patterns, and suggests that timing of activity relative to sampling behaviour is a potent cue that may enable accurate olfactory percepts to form quickly.     

高精度重力加速度的求法

为了进一步提高单摆测重力加速度的数据拟合精度,基于最小二乘法原理,用线性拟合法分别与指数拟合法、对数拟合法、幂函数拟合法及多项式拟合法结合使用的方法,获得了不同精度的重力加速度拟合值。计算结果表明:采用线性拟合法分别与指数拟合法、对数拟合法、幂函数拟合法结合使用时的数据拟合精度均比直接采用一次线性拟合法低;采用线性拟合法和多项式拟合法结合使用时的数据拟合精度比直接采用一次线性拟合法高;采用线性拟合法和多项式拟合法结合使用时的数据拟合精度随着多项式拟合次数的增加先提高后降低,出现了龙格现象。     

基于神经网络的时序预测模型研究

将神经网络作为传统的时序线性模型的非线性推广进行了分析,论证了多层前向神经网络与非线性自回归模型及反馈神经网络与非线性自回归移动平均模型的等价意义,提出了一种可作为非线性时序模型的内反馈神经网络．     

基于神经网络的复杂过程系统模型辨识

针对复杂生产过程中的一阶和二阶液位系统,利用MATLAB软件的神经网络工具箱,分别应用BP和径向基两种神经网络模型进行系统辨识,得到系统模型．通过结果比较,得出两种神经网络的应用特点：对于一阶非线性液位过程,径向基神经网络创建的数学模型性能较好;对于二阶线性液位过程,BP神经网络的建模效果较好;尽管BP神经网络的模型训练过程有学习收敛慢、局部最小点、层数和单元数不易确定的缺点,但其函数逼近的精确度对二阶线性的辨识具有独特优势．     

基于变量选择-神经网络模型的复杂路网短时交通流预测

摘要： 针对传统交通流预测模型正在由单断面历史数据处理向多断面、多时刻历史数据处理转变,但在考虑各断面间的影响时,多变的交通状况往往会使预测模型复杂化的问题,引入一种多元线性回归最小绝对收缩和选择算子方法(Lasso),并利用其优秀的变量选择能力,在复杂路网多断面中选出相关性较高的断面；结合神经网络(NN)的非线性特性,提出了Lasso NN组合模型.结果表明：Lasso NN模型在路网交叉口对未来15 min交通流数据预测的误差率低于9.2%；在非交叉口的误差率低于6.7%,总体优于各自单独使用得出的结果.     

基于组合信号源的非线性系统的子空间辨识

为解决对非线性采样系统的状态空间Hammerstein模型难以辨识的问题,提出了基于组合信号源的辨识方法.首先用组合信号源将静态非线性环节和动态线性环节分离.其次,采用模糊神经模型拟合静态非线性环节,有效地避免了采用多项式方法逼近非线性函数的限制,拓宽了非线性模型的适用范围;采用子空间算法估计采样系统的状态空间参数矩阵.最后,通过对两个非线性Hammerstein系统模型的仿真,验证了所提出的辨识方法,既简化了辨识过程,对非线性模块能够较好地拟合,又可以很快估计出状态空间方程系数矩阵,从而证明了所提方法的准确性和有效性.