Self-organized optimization of synaptic heterogeneity in recurrent neuronal networks

生物突触传递的信号受其传输特性的影响，如随时间变化的适应传输效率，最终导致突触执行时间滤波和计算。此外，这些突触传输特性是不同的--也就是说，每个神经元和突触的传输特性是不同的。然而，突触异质性是否在神经元网络的计算中发挥作用，如果是的话，网络是否能够以自组织的方式塑造突触异质性，这一事实经常被忽视，而且在很大程度上是未知的。另一方面，最近的实验表明，生物网络通过创建和移除突触不断地重新连接。此外，这一过程的功能相关性，即结构可塑性，在很大程度上尚不清楚。这个项目的主要假设是，结构可塑性提供了一种自组织机制，优化了不同种类的突触属性，从而改善了神经元网络中的时间计算。由于大脑能够在依赖于相同(感觉)输入的任务之间快速切换，我们进一步假设，这种优化(至少部分)与任务无关，而是使网络适应，为对其接收的输入进行计算提供最广泛的基础。沿着这条线，这个项目中考虑的每一个异质性特征--短期可塑性特征和突触可塑性特征--作用于不同的时间尺度，因此，可以适应输入传达信息的不同时间尺度制度。因此，我们假设结构可塑性优化了作用于不同时间尺度上的多种异质属性，可以使网络适应许多生物相关知觉任务中发生的多时间尺度输入，如语言或音乐处理或理解观察到的动作序列。我们将使用数学模型和自上而下的方法来解决这一假设。为此，我们将推导出这样一种结构可塑性机制，并在模拟中证明它允许神经元网络解决需要在多个时间尺度上整合信息的复杂感知任务。然后，将得到的算法和产生的网络与实验观察进行比较，并用来进行预测，以实验验证我们的假设。

Signals transmitted by biological synapses are influenced by their transmission characteristics, such as the adaptation transmission efficacy over time, which ultimately results in synapses performing temporal filtering and computation. These synaptic transmission characteristics are moreover heterogeneous - that is, they are different for each neuron and synapse. This fact, however, is often disregarded and it is largely unknown, whether synaptic heterogeneity plays a functional role in computation of neuronal networks and, if so, whether networks can shape synaptic heterogeneity in a self-organized manner. On the other hand, recent experiments show that biological networks are continuously rewired by creating and removing synapses. Also the functional relevance of this process, which is referred to as structural plasticity, is largely unknown. The main hypothesis of this project is that structural plasticity provides a self-organization mechanism that optimizes the heterogeneous synaptic properties to improve temporal computation in neuronal networks. As the brain is capable of rapid switching between tasks relying on the same (sensory) inputs, we further hypothesize that this optimization is (at least partly) task-independent, but rather adapts the network to provide a broadest basis for computations on the inputs it receives. Along this line, each of the heterogeneous characteristics considered in this project - short-term plasticity characteristics and synaptic plasticity characteristics - acts on a different timescale and, thus, can adapt to a different regime of timescales at which inputs convey information. Consequently, we assume that structural plasticity optimizing multiple heterogeneous properties acting on different timescales can adapt the network to multi-timescale inputs which occur in many biologically relevant perception tasks, such as language or music processing or understanding observed action sequences. We will address this hypothesis using mathematical models and a top-down-approach. To this end, we will derive such a structural plasticity mechanism and demonstrate in simulations that it allows neuronal networks to solve complex perception tasks that require integration of information over multiple timescales. The derived algorithms and resulting networks will then be compared to experimental observations and used to make predictions to experimentally test our hypothesis.