Replay in Biological and Artificial Neural Networks

大脑不断地形成对我们所经历的地方、人和事件的记忆。大脑中一个叫做海马体的区域对这一过程尤其重要--没有它，我们就不能形成对经历的记忆。然而，大脑到底是如何创建和存储这些记忆的，目前还不清楚。研究表明，记忆是以分布式神经代码的形式存储的：特定神经元的活动对应于特定的位置或信息。然而，稳定这些神经编码以供长期储存的机制尚不清楚。一种可能性是在休息时重新激活或“重放”海马神经代码。这一重播被认为加强了在海马区和分布的新皮质区域编码记忆的神经元之间的联系。然而，这种记忆重播对皮质神经回路的影响尚不清楚。我的项目的第一部分将解决海马体重播如何影响稳定皮质记忆的生理机制。最近的一项研究声称，已经使用功能磁共振成像(FMRI)记录了人类海马区的回放--这是一种间接的、非侵入性的推断大脑活动的方法。我们的研究将建立在这种方法的基础上，使用功能磁共振成像来研究海马区重演与人类新皮质记忆存储之间的关系。我的项目的第二部分将测试重放有助于从过去的经验中学习新信息的理论。有人建议，学习或抽象的信息可以用来更有效地解决问题。我们将通过再次使用功能磁共振成像来测量人类海马体的回放来测试这种假体。将重放等记忆体系结构融入到机器学习系统中，以前已经提高了它们解决任务和分析数据的能力。该项目的最后部分旨在将重放的实验结果应用于计算学习系统。

The brain constantly forms memories of the places, people and events we experience. An area of thebrain called the hippocampus is especially important for this process - without it we cannot formmemories of our experiences. However, exactly how the brain creates and stores these memoriesremains unknown. Research has shown that memories are stored as distributed neural codes: activityin specific neurons corresponds to specific places or information. However, the mechanism thatstabilises these neural codes for long term storage is unclear. One possibility is the reactivation, or'replay', of hippocampal neural codes during rest. This replay is thought to strengthen connectionsbetween neurons encoding the memory in both the hippocampus and distributed neocortical regions.However, the impact of such memory replay on cortical neural circuits is unclear. The first part of myproject will address how hippocampal replay impacts physiological mechanisms that stabilise corticalmemories. A recent study claims to have recorded replay in the human hippocampus using functionalmagnetic resonance imaging (fMRI) - an indirect, non-invasive method for inferring brain activity. Ourstudy will build on this approach to use fMRI to investigate the relationship between hippocampalreplay and memory storage in the human neocortex. The second part of my project will test theoriesthat replay facilitates learning of new information from past experience. It has been suggested thatthe learnt or abstracted information can be used to solve problems more efficiently. We will test thishypothesis by again using fMRI to measure replay in the human hippocampus. Incorporating memoryarchitectures, such as replay, into machine learning systems has previously improved their ability tosolve tasks and analyse data. The final part of the project aims to apply experimental findings onreplay to computational learning systems.