The Spiking Tolman-Eichenbaum Machine: Emergent Spatial and Temporal Coding through Spiking Network Dynamics

Abstract

The hippocampal-entorhinal system supports spatial navigation and memory by orchestrating the interaction between grid cells and place cells. While various models have reproduced these patterns, many rely on predefined connectivity or fixed weights and lack mechanisms for learning or biologically realistic temporal dynamics. The Tolman-Eichenbaum Machine (TEM) has recently gained attention as a unified generative model that explains the emergence of both grid and place cells through learning. However, existing TEM implementations rely on rate-based units and simplified architectures, which limit their biological plausibility. Here, we introduce the Spiking Tolman-Eichenbaum Machine (Spiking TEM) -- a spiking neural network model that extends the original TEM with spike-based computation and an anatomically inspired hippocampal-entorhinal architecture. Our model learns grid-like codes in the entorhinal module and context-specific place codes in the hippocampal module, while also exhibiting key temporal coding phenomena observed in electrophysiological recordings, including phase locking of spikes to theta oscillations and phase precession. Furthermore, the model gives rise to predictive grid cells in layer III of the entorhinal cortex, which prospectively encode upcoming spatial positions. These results demonstrate that structured spatial representations and temporally precise coding schemes can emerge from biologically plausible spike-based learning and dynamics, offering a unified framework for understanding spatial and temporal coding in the hippocampal-entorhinal circuit. Author SummaryAnimals, including humans, rely on an internal map of the environment to navigate and remember places. In the brain, this ability depends on two key types of neurons: place cells in the hippocampus, which fire when an animal is in a specific location, and grid cells in the entorhinal cortex, which form a hexagonal coordinate-like map of space. Understanding how neural circuits generate these spatial representations is a major goal in neuroscience. In this study, we developed a biologically grounded spiking neural network model -- the Spiking Tolman-Eichenbaum Machine (Spiking TEM)--that can learn both grid and place cell patterns through realistic neural dynamics. Our model reproduces key temporal features observed in the brain, including phase precession. In this phenomenon, neurons fire slightly earlier in each successive theta cycle as an animal moves through a place field. The model also predicts grid cells that anticipate future positions. These results provide new insight into how the hippocampal-entorhinal circuit generates and organizes spatial and memory-related representations through learning and temporal coding.