Probabilistic Learning by Rodent Grid Cells

Mounting evidence shows mammalian brains are probabilistic computers, but the specific cells involved remain elusive. Parallel research suggests that grid cells of the mammalian hippocampal formation are fundamental to spatial cognition but their diverse response properties still defy explanation. No plausible model exists which explains stable grids in darkness for twenty minutes or longer, despite being one of the first results ever published on grid cells. Similarly, no current explanation can tie together grid fragmentation and grid rescaling, which show very different forms of flexibility in grid responses when the environment is varied. Other properties such as attractor dynamics and grid anisotropy seem to be at odds with one another unless additional properties are assumed such as a varying velocity gain. Modelling efforts have largely ignored the breadth of response patterns, while also failing to account for the disastrous effects of sensory noise during spatial learning and recall, especially in darkness. Here, published electrophysiological evidence from a range of experiments are reinterpreted using a novel probabilistic learning model, which shows that grid cell responses are accurately predicted by a probabilistic learning process. Diverse response properties of probabilistic grid cells are statistically indistinguishable from rat grid cells across key manipulations. A simple coherent set of probabilistic computations explains stable grid fields in darkness, partial grid rescaling in resized arenas, low-dimensional attractor grid cell dynamics, and grid fragmentation in hairpin mazes. The same computations also reconcile oscillatory dynamics at the single cell level with attractor dynamics at the cell ensemble level. Additionally, a clear functional role for boundary cells is proposed for spatial learning. These findings provide a parsimonious and unified explanation of grid cell function, and implicate grid cells as an accessible neuronal population readout of a set of probabilistic spatial computations.越来越多的证据表明，哺乳动物大脑是概率计算机（probabilistic computers），但其中参与相关计算的具体细胞仍未明确。平行开展的相关研究指出，哺乳动物海马结构（hippocampal formation）内的网格细胞（grid cells）是空间认知的基础，但其多样的响应特性仍难以得到合理解释。尽管稳定网格现象是网格细胞领域最早发表的研究结果之一，但目前仍无合理模型能够解释动物在黑暗环境中可维持20分钟及以上的稳定网格活动。类似地，当前尚无理论能够同时解释网格破碎（grid fragmentation）与网格重缩放（grid rescaling）现象——这两种现象在环境改变时，会让网格细胞响应呈现截然不同的灵活性表现。诸如吸引子动力学（attractor dynamics）与网格各向异性（grid anisotropy）等其他特性，往往彼此矛盾，除非额外假设可变速度增益（velocity gain）等机制方能调和。现有建模研究大多忽略了网格细胞响应模式的多样性，同时也未能解释空间学习与记忆过程中，尤其是在黑暗环境下，感觉噪声（sensory noise）带来的灾难性影响。本研究通过全新的概率学习模型（probabilistic learning model），对多项已发表的电生理实验证据进行了重新解读，结果表明网格细胞的响应特性可通过概率学习过程得到精准预测。在关键操控实验中，概率型网格细胞的多样响应特性与大鼠网格细胞的实测结果在统计学上无显著差异。一套简洁自洽的概率计算框架，既能够解释黑暗环境下的稳定网格野现象、缩放竞技场中的部分网格重缩放现象、低维吸引子网格细胞动力学，以及发夹迷宫（hairpin mazes）中的网格破碎现象；同时还能将单细胞水平的振荡动力学（oscillatory dynamics）与细胞集群（cell ensemble）水平的吸引子动力学统一起来。此外，本研究还为边界细胞（boundary cells）在空间学习中的功能提出了明确的作用机制。上述发现为网格细胞的功能提供了简洁严谨且统一的解释，并表明网格细胞可作为一种可获取的神经元群体读出（neuronal population readout），反映一套概率型空间计算过程。