Information and Systems Sciences Laboratory, HRL Laboratories, LLC, Center for Neural and Emergent Systems Malibu, CA, USA.
Department of Mathematics, Center for Adaptive Systems, Graduate Program in Cognitive and Neural Systems, Center for Computational Neuroscience and Neural Technology, Boston University Boston, MA, USA.
Front Hum Neurosci. 2014 Jun 3;8:337. doi: 10.3389/fnhum.2014.00337. eCollection 2014.
The entorhinal-hippocampal system plays a crucial role in spatial cognition and navigation. Since the discovery of grid cells in layer II of medial entorhinal cortex (MEC), several types of models have been proposed to explain their development and operation; namely, continuous attractor network models, oscillatory interference models, and self-organizing map (SOM) models. Recent experiments revealing the in vivo intracellular signatures of grid cells (Domnisoru et al., 2013; Schmidt-Heiber and Hausser, 2013), the primarily inhibitory recurrent connectivity of grid cells (Couey et al., 2013; Pastoll et al., 2013), and the topographic organization of grid cells within anatomically overlapping modules of multiple spatial scales along the dorsoventral axis of MEC (Stensola et al., 2012) provide strong constraints and challenges to existing grid cell models. This article provides a computational explanation for how MEC cells can emerge through learning with grid cell properties in modular structures. Within this SOM model, grid cells with different rates of temporal integration learn modular properties with different spatial scales. Model grid cells learn in response to inputs from multiple scales of directionally-selective stripe cells (Krupic et al., 2012; Mhatre et al., 2012) that perform path integration of the linear velocities that are experienced during navigation. Slower rates of grid cell temporal integration support learned associations with stripe cells of larger scales. The explanatory and predictive capabilities of the three types of grid cell models are comparatively analyzed in light of recent data to illustrate how the SOM model overcomes problems that other types of models have not yet handled.
内嗅皮层-海马体系统在空间认知和导航中起着至关重要的作用。自从在中间内嗅皮层（MEC）的 II 层发现网格细胞以来，已经提出了几种类型的模型来解释它们的发展和运作；即连续吸引子网络模型、振荡干扰模型和自组织映射（SOM）模型。最近的实验揭示了网格细胞的体内细胞内特征（Domnisoru 等人，2013 年；Schmidt-Heiber 和 Hausser，2013 年）、网格细胞的主要抑制性递归连接（Couey 等人，2013 年；Pastoll 等人，2013 年）以及网格细胞在 MEC 背腹轴上多个空间尺度的解剖重叠模块内的拓扑组织（Stensola 等人，2012 年），为现有网格细胞模型提供了强有力的约束和挑战。本文提供了一个计算解释，说明 MEC 细胞如何通过在模块化结构中具有网格细胞特性的学习而出现。在这个 SOM 模型中，具有不同时间整合率的网格细胞会学习具有不同空间尺度的模块化特性。模型网格细胞会根据来自多个方向选择性条纹细胞的输入进行学习（Krupic 等人，2012 年；Mhatre 等人，2012 年），这些细胞对在导航过程中经历的线性速度进行路径整合。较慢的网格细胞时间整合率支持与较大尺度的条纹细胞建立学习关联。根据最近的数据，比较分析了这三种类型的网格细胞模型的解释和预测能力，以说明 SOM 模型如何克服其他类型的模型尚未解决的问题。
