预印本：预测性网格细胞编码不依赖显式规划

Predictive grid cells, a recently discovered functional class of neurons in medial entorhinal cortex (MEC), exhibit grid-like tuning, not with respect to where an animal currently is, but where it will be. Such coding, which is disjoint from standard grid cell coding, appears consistent with the hypothesis that MEC supports predictive computations for goal-directed navigation. However, given that existing predictive coding theories have been used to understand standard grid cells, it remains unclear what the functional significance of predictive grid cells is. Furthermore, the extent to which predictive grid cells are generally present in MEC is unknown. To address these open questions, we analyze data from rodent electrophysiology recordings and recurrent neural networks (RNNs). In an independent, previously recorded MEC dataset, we find evidence of predictive grid cell coding when the animal is awake and exploring an open field environment, suggesting that they are a general feature of rodent MEC. In RNNs trained to path-integrate, which have previously been shown to develop units with grid-like representations, we also find units with predictive grid-like representations, even though the RNNs are not trained with a predictive objective. We validate the predictive grid tuning of these units by finding that they exhibit significant future path-dependence, even when controlling for the same allocentric spatial position. In addition, we find that band-like units, which have been found to play an important role in maintaining path-integration in the RNNs, exhibit greater predictive grid tuning than expected by chance. These results lead us to hypothesize that predictive grid cells are a property of the underlying grid cell circuit. We test this by looking for signatures of predictive grid coding in the MEC electrophysiological data collected during sleep, finding evidence in one recording. Collectively, our results challenge the idea that predictive grid cells must ne

中文翻译还没有生成，先显示原文。