Oxford Centre for Computational Neuroscience, Oxford, UK.
Department of Computer Science, University of Warwick, Coventry, UK.
Hippocampus. 2024 Nov;34(11):608-624. doi: 10.1002/hipo.23636. Epub 2024 Sep 2.
A key question for understanding the function of the hippocampus in memory is how information is recalled from the hippocampus to the neocortex. This was investigated in a neuronal network model of the hippocampal system in which "What" and "Where" neuronal firing rate vectors were applied to separate neocortical modules, which then activated entorhinal cortex "What" and "Where" modules, then the dentate gyrus, then CA3, then CA1, then the entorhinal cortex, and then the backprojections to the neocortex. A rate model showed that the whole system could be trained to recall "Where" in the neocortex from "What" applied as a retrieval cue to the neocortex, and could in principle be trained up towards the theoretical capacity determined largely by the number of synapses onto any one neuron divided by the sparseness of the representation. The trained synaptic weights were then imported into an integrate-and-fire simulation of the same architecture, which showed that the time from presenting a retrieval cue to a neocortex module to recall the whole memory in the neocortex is approximately 100 ms. This is sufficiently fast for the backprojection synapses to be trained onto the still active neocortical neurons during storage of the episodic memory, and this is needed for recall to operate correctly to the neocortex. These simulations also showed that the long loop neocortex-hippocampus-neocortex that operates continuously in time may contribute to complete recall in the neocortex; but that this positive feedback long loop makes the whole dynamical system inherently liable to a pathological increase in neuronal activity. Important factors that contributed to stability included increased inhibition in CA3 and CA1 to keep the firing rates low; and temporal adaptation of the neuronal firing and of active synapses, which are proposed to make an important contribution to stabilizing runaway excitation in cortical circuits in the brain.
理解海马体在记忆中的功能的一个关键问题是如何将信息从海马体回忆到新皮层。这在海马体系统的神经元网络模型中进行了研究，其中“什么”和“哪里”神经元发放率向量被应用于分离新皮层模块，然后激活内嗅皮层的“什么”和“哪里”模块，然后是齿状回，然后是 CA3，然后是 CA1，然后是内嗅皮层，然后是向新皮层的逆行投射。一个速率模型表明，整个系统可以被训练为从新皮层中的“什么”应用作为检索线索来回忆“哪里”，并且原则上可以朝着由任何一个神经元上的突触数量除以表示的稀疏性决定的理论容量进行训练。然后将训练后的突触权重导入到相同架构的整合和发射模拟中，该模拟表明，从向新皮层模块呈现检索线索到回忆整个新皮层记忆的时间约为 100 毫秒。这对于逆行投射突触在情节记忆存储期间被训练到仍处于活动状态的新皮层神经元来说足够快，并且这对于正确地向新皮层进行回忆是必要的。这些模拟还表明，在时间上连续运行的长环路新皮层-海马体-新皮层可能有助于在新皮层中进行完整的回忆；但是这种正反馈长环路使整个动力系统固有地容易受到神经元活动病理性增加的影响。有助于稳定性的重要因素包括 CA3 和 CA1 中的抑制增加，以保持低发放率；以及神经元发放和活性突触的时间适应，据提议，这对稳定大脑皮质电路中的失控兴奋做出了重要贡献。
