Abstract
Hippocampal place cells exhibit varying degrees of directionality depending on environmental constraints. Experimental data shows that place cells are mostly directional in radial or plus-mazes, whereas they are largely non-directional in open environments. We hypothesize that place cells are initially directional, responding to specific local views, and that this directionality is reduced through exploration when an agent is not constrained in its movements. We propose a feed-forward model of vision-driven place cells that correlates local views for different headings at the same location. The model is evaluated using both idealized landmark data and real-world visual stimuli processed via Gabor filters, demonstrating that exploration and Hebbian learning are sufficient to produce place cells with reduced directionality in open environments.
Introduction
Place cells in the hippocampus are renowned for their ability to fire when an animal occupies a specific spatial location. However, their firing properties are not static; they are highly directional in constrained environments like mazes but become omnidirectional in open fields. This paper investigates whether a simple feed-forward neural mechanism, rather than complex recurrent dynamics, can explain this transition. We explore the role of visual input and exploration history in shaping the spatial and directional characteristics of these cells.
Model Architecture
The proposed model utilizes a two-layer feed-forward architecture consisting of View Cells (VCs) and Place Cells (PCs). VCs represent the input layer, encoding sensory information from the environment. These cells can represent either idealized landmarks (distance and bearing) or features extracted from real-world panoramic images.
The connections between VCs and PCs are governed by weights $w_{ij}$, which are updated using unsupervised Hebbian learning. As an agent explores, the PC learns to associate multiple VCs—each representing a different view of the same spot—into a single spatial representation. In an open field, where an agent can rotate 360 degrees at any point, the PC eventually pools all directional views, resulting in a non-directional place field.
Visual Processing and Environments
To test the model's robustness, the authors implemented a Gabor-based visual processing system. Panoramic images from real-world environments were processed using Gabor filters to simulate biological visual feature extraction. This approach introduces "perceptual aliasing," where different locations may appear visually similar, challenging the model to maintain distinct place fields.
The model was evaluated in two simulated settings: an 80x80cm open field and a plus-maze with 40cm arms. A Directionality Index ($D$) was used to quantify the results, where 1 indicates perfect directionality and 0 indicates perfect omnidirectionality.
Results and Discussion
The landmark-based model successfully replicated experimental findings. In the plus-maze, where movement is restricted, the directionality index remained high ($D \approx 0.84$). Conversely, in the open field, $D$ dropped significantly to $0.16$. The Gabor-based model followed the same trend, although it exhibited higher overall directionality ($D \approx 0.62$ in open fields) due to the complexity and noise of real visual data.
These results suggest that the transition from directional to non-directional firing is a natural consequence of how freely an agent can explore. In constrained corridors, the agent rarely sees the "reverse" view of a location, preventing the Hebbian pooling necessary to reduce directionality.
Conclusion
This research demonstrates that a vision-driven feed-forward architecture is sufficient to explain the emergence of non-directional place fields in open environments. The primary driver of this transition is the agent's ability to view the same location from multiple headings during exploration. Future work will focus on integrating path integration to mitigate perceptual aliasing and incorporating multisensory data to create more robust spatial representations.
Summary updated