Neural coding of visual inputs in the human medial temporal lobe

视网膜捕捉到的图像由神经元传输到枕叶的主要视觉区域。经过几个处理阶段后，这些信息被传递到下颞叶皮质，并从那里投射到内侧颞叶(MTL)。正是在这里，PI一直在与加州大学洛杉矶分校医学中心的Itzhak Fry实验室合作，记录单个神经元对视觉刺激的反应。这些记录来自患有癫痫的人类患者，这些患者被植入颅内电极，以确定癫痫发作的病灶，以便进行可能的手术切除。最近，有研究表明，很大比例的MTL神经元对名人、熟悉的个人、地标和动物的照片做出反应(Quian Quiroga等人，自然，435：1102-1107；2005)。这种反应是高度选择性的，因为这些神经元中的每一个只对极少数的图片做出反应。与分布式代码不同，在分布式代码中，信息由大量细胞中放电活动的特定分布隐式表示，这些数据与MTL中存在的显式表示相一致。换句话说，个体或物体的身份是由少数神经元代表的，而这些神经元中的每一个都明确地告诉我们所看到的东西的身份。一群神经元代表了多少信息，可以通过根据相关神经元的反应模式重建刺激来客观地量化。这种重建或解码为理解大脑如何提取特征并破译在神经元群体活动中编码的信息提供了一种有价值的方法。考虑到在MTL神经元中发现的显式表征，为了进一步了解知觉在MTL中的表征方式，我们计划研究是否有可能从同时记录的MTL神经元的活动中预测在每一次试验中显示了哪幅图。这样的研究直接解决了一个更普遍的问题，即大脑中的神经元如何代表人类的视觉信息。除了执行常规分析外，还建议开发新的方法来分析这些数据。植入电极附近神经元的活动反映在棘波上。每个电极检测其周围所有神经元的尖峰，每个神经元发射给定形状的尖峰。为了了解这些单位如何在大脑中编码信息，必须有一种可靠的方法来检测和分类棘波(即根据其形状将棘波从不同的神经元中分离出来)。PI已经开发了一种用于检测和分类棘波的自动方法(Quian Quiroga等人，Neural Computation，16：1661-1687；2004)，并计划进一步改进该算法。特别是，通过更精细的参数调整，有可能以自动的方式获得更多的神经元。此外，算法的进一步优化可以使其运行更快，甚至在线。预计这些改进将对其他动物神经生理学实验室有用。这对于其他应用尤其重要，因为自动分拣是强制性的，例如开发脑机接口和神经假体(即由神经信号直接引导的假体设备)。根据神经元的活动预测视觉输入的可能性还取决于用于解码的算法，以及这些算法读出什么类型的信息来进行预测。在这方面，研究不同译码算法的性能以及神经元反应的哪些特征对译码来说是最重要的，这是很重要的。

Images captured by the retina are transmitted by neurons to primary visual areas in the occipital lobe. After several processing stages, this information is transmitted to the inferior temporal cortex and from there it projects to the medial temporal lobe (MTL). It is from here that the PI has been recording individual neurons responding to visual stimuli in collaboration with the laboratory of Itzhak Fried at the UCLA medical center. These recordings are from human patients suffering from epilepsy, which are implanted with intracranial electrodes in order to determine the seizure focus for possible surgical resection. Very recently it has been shown that a large proportion of MTL neurons respond to pictures of celebrities, familiar individuals, landmarks, and animals (Quian Quiroga et al., Nature 435: 1102-1107; 2005). Such responses were highly selective in the sense that each of these neurons responded only to very few pictures. In contrast to distributed codes in which information is represented implicitly by the specific distribution of firing activity in large populations of cells, these data are in agreement with the existence of an explicit representation in MTL. In other words, the identity of individuals or objects is represented by a small number of neurons and each of these neurons is 'explicitly' telling the identity of what is being seen.How much information is represented by a population of neurons can be quantified in an objective manner by reconstructing the stimulus from the patterns of responses of the neurons in question. Such reconstruction, or decoding, offers a valuable approach to understand how the brain extracts features and deciphers information encoded in the activity of population of neurons. Given the explicit representation found in MTL neurons, in order to get further insights on how percepts are represented in MTL, we plan to study whether it is possible to predict in each trial which picture was shown from the activity of simultaneously recorded MTL neurons. Such a study addresses directly the more general issue of how neurons in the brain represent visual information in humans. Besides performing conventional analysis, it is proposed to develop new methods to analyze this data. The activity of neurons in the vicinity of the implanted electrodes is reflected in spikes. Each electrode detects the spikes of all neurons in its surroundings and each neuron fires spikes of a given shape. In order to understand how these units encode information in the brain, it is mandatory to have a reliable way of detecting and sorting the spikes (i.e. separating the spikes from the different neurons based on their shapes). An automatic method has been developed by the PI for detection and sorting of spikes (Quian Quiroga et al, Neural Computation, 16: 1661-1687; 2004) and it is planned to further improve this algorithm. In particular, with finer tuning of parameters it may be possible to obtain more neurons in an automatic way. Moreover, further optimization of the algorithm can make it run faster and even on-line. It is expected that these improvements will be useful to other animal neurophysiology laboratories. In particular, this is important for other applications were automatic sorting is mandatory, such as the development of Brain Machine Interfaces and Neural Prosthesis (i.e. prosthetic devices guided directly by neural signals). The possibility of predicting visual inputs from the activity of neurons also depends on the algorithm used for decoding and what type of information is read-out by these algorithms for doing the predictions. In this respect, it is important to study the performance of different decoding algorithms and also to study what features of the neuronal responses are the most important for decoding.