Models of Neural Credit Assignment Mechanisms

学习是日常生活的核心要素。导致学习的潜在神经生理过程被称为突触可塑性。突触可塑性这一术语包含了两个神经元之间通过各自突触的强度增强或减弱的通信的所有过程。这些更改可以持续几秒到几分钟的短时间范围，也可以持续几天甚至几年的长时间。突触可塑性的确切机制尚不清楚。尽管存在大量的实验文献，但关键问题仍无法得到令人满意的回答：如何将信用或奖励信号分配给单个突触，从而使神经网络的性能通过这些标记连接的适应而得到增强。这个项目将探索在尖峰神经网络中可能的算法来解决上面段落中概述的信用分配问题，尖峰神经网络是一个能够模拟大脑中发现的大多数神经元尖峰行为的神经网络家族。学习算法将根据现有的神经科学实验数据构建，以确保生物学上的合理性，并提供基于实验可验证的模型形成可测试假设的可能性。因此，该项目将有助于更好地理解大脑学习的基本机制。________________________________________

Learning is a core element of the every day life. The underlying neurophysiological processes enable learning are called synaptic plasticity. The term synaptic plasticity incorporates all processes by which the communication between two neurons via the strength of their respective synapse either get enhanced or depressed. These changes can persist for short temporal scopes in the range of seconds to minutes or long durations lasting for days or even years. The exact mechanisms of synaptic plasticity are yet to be uncovered. Even though a huge body of experimental literature exists key questions could not be answered satisfactorily: how does credit or a reward signal gets assigned to single synapses so that the neural networks performance get enhanced by the adaption of these tagged connections.This project is going to explore possible algorithms for the credit assignment problem outlined in the paragraph above in spiking neural networks, a family of neuronal networks able to model the majority of spiking behaviors of neurons found in the brain. The learning algorithms will be constructed along existing experimental data from neuroscience to ensure biological plausibility and offering the possibility to form testable hypothesis based on the model verifiable by experiments. Therefore, the project would contribute to a better understanding of the fundamental mechanisms of learning in the brain. ________________________________________