arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1505.04432q-bio.NC
收藏

结构突触可塑性形式模型中的耦合Hebbian学习和演化动力学

Coupled Hebbian learning and evolutionary dynamics in a formal model for structural synaptic plasticity

Harold P. de Vladar, Eörs Szathmáry

摘要:

神经元功能的理论模型考虑了网络学习、分类和识别输入的不同机制。这些模型的一个中心焦点是了解如何在神经元之间建立关联,以便预测与经验观察相兼容的尖峰模式。尽管这些模型带来了重大的见解和进步,但它们仍然没有解释大脑解决某些问题的惊人速度以及其创造力背后的原因。我们研究了两个重要的组成部分,可能至关重要的援助,全面了解说,神经动力学过程。首先,我们认为,一旦提出了一个问题,不同的假定的解决方案是由不同的群体或当地的神经元复合体并行产生的,随后的稳定和传播的最佳解决方案。使用数学模型,我们表明,这种机制加速找到正确的解决方案。这种形式类似于标准的进化复制-变异模型,其中变异类似于神经元状态切换(开/关)的概率。我们纳入的第二个因素是结构突触可塑性,即新突触的形成和旧突触的解体,我们将其应用为突触连接的动态重组。我们表明,赫布学习本身并不足以达到最佳的解决方案。然而,将其与并行评估和结构塑性相结合,为有效解决问题提供了可能性。在由此产生的网络中,拓扑收敛到全连接组件的子集。对突触施加成本会降低连接性,尽管连接组件的数量仍然很强大。对于早期建立的连接,突触的平均寿命更长,并且随着突触成本的减少而减少。

英文摘要:

Theoretical models of neuronal function consider different mechanisms through which networks learn, classify and discern inputs. A central focus of these models is to understand how associations are established amongst neurons, in order to predict spiking patterns that are compatible with empirical observations. Although these models have led to major insights and advances, they still do not account for the astonishing velocity with which the brain solves certain problems and what lies behind its creativity, amongst others features. We examine two important components that may crucially aid comprehensive understanding of said neurodynamical processes. First, we argue that once presented with a problem, different putative solutions are generated in parallel by different groups or local neuronal complexes, with the subsequent stabilization and spread of the best solutions. Using mathematical models we show that this mechanism accelerates finding the right solutions. This formalism is analogous to standard replicator-mutator models of evolution where mutation is analogous to the probability of neuron state switching (on/off). The second factor that we incorporate is structural synaptic plasticity, i.e. the making of new and disbanding of old synapses, which we apply as a dynamical reorganization of synaptic connections. We show that Hebbian learning alone does not suffice to reach optimal solutions. However, combining it with parallel evaluation and structural plasticity opens up possibilities for efficient problem solving. In the resulting networks, topologies converge to subsets of fully connected components. Imposing costs on synapses reduces the connectivity, although the number of connected components remains robust. The average lifetime of synapses is longer for connections that are established early, and diminishes with synaptic cost.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑