Dendritic Computation and the Neural Code

如果把扬声器和植入脑细胞的电极连接起来，就会听到一种粗糙的、噼啪作响的、明显没有结构的声音。这种噪音是否反映了信息处理有机形式的内在不精确，或者更确切地说，它是一种未知的、可能是高度优化的信息编码方式的结果？神经科学研究的核心是神经编码问题。现在已经很清楚，周围神经以它们的尖峰速率编码来自感官的信息。然而，新皮层层次结构中的神经元不断地将两种不同性质的信息组合在一起：一种是直接来自感官的自下而上的信息，另一种是来自内部来源的自上而下的信息。因此，我们提出了一个简单的神经编码问题的重新表述，并提出了一个一般性的问题：单个神经元群体如何同时编码两个信息流？最近的实验证据表明树突在回答这个问题方面起着关键作用。利用新皮层网络的数值模拟，该资助将(1)确定树突依赖性爆发在同时表示自上而下和自下而上信息方面的作用。此外，模拟将用于(2)研究抑制连接基序在优化突发神经编码中的作用。最后，我们将(3)开发统计数据分析方法，以促进树突依赖性突发编码的实验研究。当今最强大的机器学习方法——深度学习，就是受到了大脑皮层分层结构的启发。通过在层次结构中概述神经编码规则，所提出的工作可以激发信号处理算法的有效实现。此外，理解新皮层使用的神经代码对于分析生物医学数据至关重要。为了找出一个可能的应用领域，我们注意到脑机接口技术的改进在很大程度上取决于本提案中讨论的新型解码算法。因此，我们对神经编码问题的新方法可以带来有价值的技术。

If one connects loudspeakers to an electrode implanted in brain cells, one would hear a rough, crackling and apparently unstructured sound. Is this noise a reflection of intrinsic imprecisions of an organic form of information processing or, rather, is it the result of an unknown and perhaps highly optimized way of encoding information? At the center of neuroscience research lies this problem of neural coding. It has become clear that peripheral nerves code information streaming from the senses in their spiking rate. Neurons within the hierarchical structure of the neocortex, however, constantly combine information of two different natures: bottom-up information coming more directly from the senses and top-down information coming from internal sources. Therefore, we propose a simple reformulation of the neural coding problem and ask the general question: How can a single population of neurons encode two streams of information simultaneously? Recent experimental evidence point to a pivotal role of dendrites in answering this question.Using numerical simulations of neocortical networks, this grant will (1) determine the role of dendrite-dependent bursting for representing top-down and bottom-up information simultaneously. In addition, the simulations will be used to (2) investigate the role of inhibitory connection motifs to optimize the bursting neural code. Lastly, we will (3) develop statistical data analysis methods to facilitate experimental investigations of dendrite-dependent burst coding.The most powerful machine learning method of today, deep learning, was inspired by the hierarchical structure of the neocortex. By outlining the rules for neural coding in a hierarchy, the proposed work can inspire efficient implementations of signal processing algorithms. In addition, understanding the neural code used by the neocortex is essential to the analysis of biomedical data. To single out a possible area of application, we note that the improvement of brain-machine interface technology strongly depends on novel decoding algorithms of the type discussed in this proposal. Therefore, our novel approach to the problem of neural coding can lead to valuable technologies.