天上的小石子

Xu Y, Turk-Browne NB and Chun MM. (2007). Dissociating Task Performance from fMRI Repetition Attenuation in Ventral Visual Cortex. Journal of Neuroscience. 27(22) : 5981 - 5985.

This paper argued that task performance improvements commonly associated with behavioral priming can be independent with repetition attenuation. Behavioral priming means increased response accuracy and speed with repeated presentation of similar stimuli. Repetition attenuation means reduced neural responses elicited by repeated visual stimuli. It is intuitive to think that repetition attenuation merely reflects decreased processing time, which in turns results in priming. But in this paper the authors conducted two exepriemtns and found that this is not true. In the scene task the participants were asked to judge whether two pictures came from the same scene; in the image task, they were asked to judge whether two images are identical. While the participants’ improvements were found at more similar pictures in the scene task and at less similar pictures in the image task, repetition attenuation were identified by fMRI in the parahippocampal place area (PPA) at more similar stimuli for both tasks. Then the authors argued that attenuation in ventral visual areas reflects stimulus-specific processing independent of task demands.

This result may also suggest that at least in the ventral visual pathway, reduced neuronal responses may be only related to stimulus properties (e.g., repeated presentation), rather than higher-order decision making circuits. The mechanisms underlying priming may be more complicated than repetition attenuation.

Hinton G.E.* and Salakhutdinov R.R. (2006). Reducing the Dimensionality of Data with Neural Networks. Science Vol 313 : 504 - 507.

Preview:
Cottrell G.W. (2006). New Life for Neural Networks. Science Vol 313 : 454 - 455.

This paper describes a new method for dimension reduction in large datasets. The authors used a multilayered ‘autoencoder’ neural network to represent highly nonlinear mappings from higher dimensions to lower dimensions, and from lower dimension back out to higher dimensions. This way, a reduced representation of complex dataset is found. It occurs to me that such reduced representation may work in a way comparable with neural encoding in the brain visual system, which ‘compresses’ the complex stimuli into concise representations, which are then in turn restored to some extent when ‘read out’ by other functionalities in the brain. This kind of representation should be optimized to allow efficient data restoration, and therefore it should provide most amount of information. Indeed, in the brain the restored data is not necessarily the same as (or approaching) the input, as the brain may depoly different approaches to utilize it.

Jaramillo S and Pearlmutter BA. (2007). Optimal Coding Predicts Attentional Modulation of Activity in Neural Systems. Neural Computation 19, 1295–1312.

This work demonstrates a simple feedforward network capable of selective attention to input visual stimuli. The attention modulation signal is represented as a two-dimensional vector indicating the position of attention center in the input space. During the training session, a cost function weighted by attentional importance is minimised in order to achieve an optimized network (in this work, so that the decoded output is same as the input layer). This work shows that attentional modulation is not, at its root, due to specific local architectural features but is rather a ubiquitous phenomenon to be expected in any system with shifting fidelity requirements (i.e. changing attentional signal).

It occurs to me that this simple attentional modulation mechanism can be easily implemented in Topographica. A new sheet thus receives attention signals and directs a ‘attentional importance’ topography as additional inputs to feedforward sheets. However it is not clear how the information-theoretic accounts can be exactly implemented in Topographica, but the self-organizing network should generally reduce overall cost if a desired map is formed.

Herzog A. et. al. (2007). Displaced strategies optimize connectivity in neocortical networks. Neurocomputing 70:1121–1129.

This paper postulates a novel neocortical connection method, termed displaced connection strategy. This method forms the connection cluster centers at a distance of the source neuron, either locally or globally. This way, the authors mentioned that minimal wiring cost is achieved. This method is also biologically plausible and can produce sparse connection network.

Disadvantages of other connection method:
random local connections: hard to achieve synchronization between neurons;
global random connections and small-world models: not biologically plausible.

An inspiration to Topographica from this approach is the new mechanism of lateral connections, which may adopt this displaced strategy to form recurrent (lateral) connections (e.g., displaced long-term inhibition v.s. local short-term excitation).

Kai-Markus Müller. (2007). Acquiring Visual Object Expertise: Reorganization in the Ventral Path. The Journal of Neuroscience 27(17):4497-4498.

This journal club paper reviewed:

Hans P. Op de Beeck HP. et. al. (2006). Discrimination Training Alters Object Representations in Human Extrastriate Cortex. The Journal of Neuroscience, 26(50):13025-13036.

In this paper, high resolution fMRI evidences were shown to the reorganization of object representation after discrimination training in human lateral occipital complex (LOC). Subjects were asked to discriminate among three categories of objects before and after training. During training, one subject became expert of one category of objects. Two predictions were made in terms of BOLD responses before and after training. 1) The voxel activity of pre-training scan would highly correlate with the post-training scan, along with a topographically related enhancement. 2) “Rewiring” of object-related areas might result in a rather unrelated increase in activity, and thus the topographical pattern of activity for trained stimuli would be less correlated than for untrained stimuli.

Experiment results showed that correlation between two scans were decreased in the trained object case than untrained case. This speaks in favor of a reorganization of object related areas rather than a mere intensity enhancement of previously activated voxels. In the modelling perspective, this means that training of certain type of objects should result in first increased activity and then reorganization topography. It can also be assumed (or predicted) that between increase and reorganization, there should an ‘adaptation (decrease)’ process if certain types of objects are continuously presented.

转载自“格志”：

PhysicsWeb新闻上是这么说的：光照射一个叶绿素分子，分子上的电子被激发。一般人们都认为，激发的分子会随机的把能量传递到最近的能量比较低的分子上，这个过程就是所谓的“downhill”。能量会不断的沿着下坡路往下跑，直到达到发生光合反应的中心。

问题在于，这个能量随机传递的模型无法解释光合作用效率为什么这么高：95%以上。在这个实验中，人们发现，能量的传递并不是这种经典的随机过程，而是一个量子运动过程。他们发现叶绿素分子被激发后，震荡信号维持了几百飞秒，这被认为是出现了量子拍(quantum beats)，也就是说所有的能级同时相干的被联系到一起了。换句话说，激发态同时感觉到了所有的这些能态，但并没有真的访问这些能态，于是它可以通过最优的路径达到反应中心，而不损失能量。

这个过程，实际上就与1997年发明的Grover量子搜索算法思想上很类似。“在这个光合反应中，能量搜索它需要追随的路径，到达它发挥效用的地方。”密歇根大学的Roseanne Sension解释说。

争议仍旧存在，因为实验是在77K的低温下进行的，所以对处于300K的常温下的树叶来说，是否也会出现这个效应，人们不能肯定。

量子现象在分子级或以下尺度广泛存在，不过这大概是第一次将量子效应同某种生物活动联系起来。在叶绿素分子上进行量子搜索产生了伟大的光合作用，那么神经细胞间及其快速的信号处理是否也可能同量子效应相关？著名的（同时也是广受争议的）Penrose-Hameroff模型（Penrose，罗杰·彭罗斯，比较有名的科普作品有《皇帝新脑》、《时间之箭》等）：Orchestrated Objective Reduction of Quantum Coherence in Brain Microtubules: The “Orch OR” Model for Consciousness（大脑微管中协调的量子相干态自塌缩：意识的”Orch OR”模型）就是关于大脑和量子效应的猜想：

神经细胞可能远比0和1的开关要复杂。如果我们观察神经细胞的内部，就会发现由微管和其它纤维状结构组成的高度有序的网络（细胞骨架），它们共同作用产生细胞活动。微管是圆柱状微管蛋白聚合物，组织成六角形栅格构成的柱体壁。
微管中微管蛋白间的相互作用可能同信息处理有关。而因为微管蛋白的状态是由量子内力（范德瓦耳斯力）控制，它们可能处在多重量子叠加状态（qubits），故而微管可以被看成细胞中的量子计算机。彭罗斯1994年提出的“波函数自塌缩”发生在微管中，导致某种微管-微观蛋白构象态的形成，这种量子态控制包括神经突触间的活动。塌缩后微观蛋白状态的概率受微管相关蛋白（MAPs）的影响，它们协调量子震荡。由这种微观蛋白构象态形成的宏观相干叠加可能在大脑中广泛存在，并同意识密切相关。

那么说如果每一个神经细胞都是一台量子计算机，大脑中10^11个神经细胞就组成了一个高度互联的量子计算机网络。这样的复杂度已经超过了人类的理解范围。

这个猜想相当玄妙，相当大胆，因而也受到了大量质疑。然而伟大的理论刚一出世往往是备受指责和争议的，就像相对论和量子论刚被提出时那样。大脑中许多不可思议之处，除了高度的复杂性之外，就是高速信号传递和同步。在神经研究中发现的神经细胞活动的同步性，和发现很久但依然神秘的脑电波信号（EEG）的同步性，大概可以说明以上量子大脑猜想的合理性。如果像高效的光合作用能在一定条件下用量子效应验证那样验证这种“微管态”的量子现象，许多未知的原理便可以有更合理的解释。

Mirabella G et. al. (2007). Neurons in Area V4 of the Macaque Translate Attended Visual Features into Behaviorally Relevant Categories. Neuron 54(2) : 303 - 318.

Preview:
Vogels R. (2007). Representation of Response Categories in Visual Cortex. Neuron 54(2) : 181 - 183.

This paper suggests that the later part of the response of V4 neurons can be modulated by the behavioral response category: the initial representation in V4 is strongly feature-driven, while at a later phase of the neural response, the behavioral category that is associated with the attended feature is also coded by the V4 neurons.

It is important to address, in future research, whether these behavioral response category-related responses originate in V4 or are due to feedback from other visual or nonvisual areas of the brain. If they are from V4, V4 neurons might play a role beyond visual analysis by encoding the stimulus features in a categorical format that is directly relevant to control the behavioral performance. But if from feedback, they do not have a causative role in the behavior of the animal.

This V4 behaviour also raises the question of whether similar modulations might be present in other visual areas, possibly even in area V1 (e.g. Shuler and Bear. (2006). Reward Timing in the Primary Visual Cortex. Science 311 : 1606 - 1609.). Intriguing enough, studies yet have failed to provide any evidence for such category-related responses in inferior temporal cortex where V4 projects.

《自然》杂志这周报道（新闻、评论和论文），来自斯坦福大学和马普生物物理所的合作组发现了两种在单细胞生物中存在的感光蛋白质ChR2和NpHR能够“开关”神经细胞的活动。科学家将它们在转基因哺乳动物的神经细胞中表达，当用蓝色照射神经细胞时，其膜上的ChR2使得细胞激活，当用黄光照射神经细胞时，其膜上的NpHR使得细胞停止活动。这一技术非常强大，因为ChR2和NpHR不影响细胞的正常活动，被激活或者抑制的细胞活动都是可逆的，反应速度非常快（大约50纳秒），并且实验者能够在远程操作。

ChR2是细胞膜上的一个阳离子通道，当它被蓝光激活时，外界的钠离子和钙离子进入细胞，内部的钾离子流出，细胞被去极化（depolarization），产生动作电位（action potential）。NpHR是细胞膜上的一个离子泵，当它被黄光激活时，外界的氯离子进入细胞，使它极化（hyperpolarization），动作电位被抑制。非常幸运的是，实验得出的光波长-激活关系图显示蓝光和黄光的分布恰巧分开，这样一个细胞同时可以表达ChR2和NpHR，而这两种蛋白质只会被一种光激活。事实上，研究者可以用连续的光谱来远程加强或减弱某种细胞的活动，也可以用通过让细胞表达其它的荧光指示剂（例如指示细胞间钙离子浓度可以反映细胞的活动），然后用蓝-黄光谱之外的光来观察细胞的活动状况。

这一技术能够为大脑神经网络的研究和神经活动-行为关系的研究带来巨大的飞跃。ChR2–NpHR系统能够非常容易地记录某种感官输入导致的神经活动模式，然后在没有输入的情况下“控制”神经细胞使得记录的模式“重放”，从而产生与先前一致的行为，或者“关闭”某些神经的活动，使得某种行为立刻停止。研究者在秀丽隐杆线虫（C. elegans）上使用ChR2–NpHR系统，成功地操纵控制肌肉活动的神经细胞，使它按照实验者的意愿停止/恢复游泳的状态。这种线虫不是哺乳动物，研究在ChR2–NpHR系统上进行了额外处理使得这个系统也能够工作。

这一技术能够淘汰传统的电极探针和化学激发/抑制剂方法，在了解大脑原理、医治神经系统疾病（例如癫痫、精神分裂症、帕金森氏症、老年痴呆症等）方面将发挥重大作用。使用光技控制神经的活动可能加强人类感知，最终完全控制我们的大脑和行为。

Cheetham CEJ, Hammond MSL, Edwards CEJ and Finnerty GT. (2007). Sensory Experience Alters Cortical Connectivity and Synaptic Function Site Specifically. The Journal of Neuroscience, 27(13):3456 –3465

In adulthood rats (i.e. the whisker cortical maps have developed) with trimmed whiskers, synaptic inputs to layer 2/3 pyramidal neurons were altered at the junction of deprived and spared cortex in primary somatosensory cortex. There is a reduction of local excitatory connectivity in deprived cortex, but no compensatory increases in the strength of remaining local excitatory connections; remarkably, there is a strengthening of local excitatory connections in spared cortex, without change of local excitatory connectivity and synaptic number.

Possible model: In freely behaving rats, sensory deprivation is followed acutely by increased firing in spared cortex and decreased firing in deprived cortex. The increased firing in the spared cortex has two effects. 1), it results in strengthening of existing excitatory synapses in spared cortex, which manifests as potentiation of local excitatory connections. 2), increased firing or changes in the pattern of firing in spared cortex will augment the excitatory drive from spared to deprived cortex. And this promotes loss of excitatory synapses in deprived cortex. Hence, depression of sensory responses in deprived cortex and potentiation of sensory responses in spared cortex can be dissociated by manipulations that block synaptic potentiation but do not affect the acute changes in firing.

Michel, M. M., & Jacobs, R. A. (2007). Parameter learning but not structure learning: A Bayesian network model of constraints on early perceptual learning. Journal of Vision, 7(1):4, 1–18.

Learning processes underlying cue acquisition are biased by prior beliefs about potentially dependent variables such that cue acquisition is possible when a signal is correlated with a cue to a scene property and the signal is potentially dependent on that property (e.g. novel auditory signal as a cue to motion direction). If the signal is not believed to be potentially dependent on the property (e.g., binocular disparity or brightness signals as cues to motion direction), cue acquisition fails.

Refer the former as ‘parametric learning’ and latter as ’structural learning’. This means that the conditional probability P(cue|property) is different a prior for different cues. Already established relation (parametric learning) can induce correct posterior P(property|cue) quickly, while naturally independent relation (structural learning) needs much extra learning to modify the conditional probability.

This may suggest in low level, after learning there are some conditional probability between input stimuli (cue) and a neuron’s (or a population of neurons’) response (property). Such probability as a whole may show the correlation of the input-response space. Through extensive learning such dependency can be altered. In a LISSOM map, we can measure conditional probabilities through running the simulation and calculating the responses. One thing we can do is to check if the input-output reflect the correct correlation (maybe add some noise?), another thing we can record a probability history from a established map, and use it to test or modify the new inputs.