When navigating a new environment, we learn its layout through two primary streams of information: spatial contiguity (things physically close to each other) and temporal contiguity (things encountered one after another). Because we move continuously, these two streams usually align. But what happens when they don't? Do we build our mental maps based on the sequence of our experiences, or do we rely on hard spatial coordinates?
This paper explores this long-standing psychological debate using neural networks as a "computational laboratory," investigating how these two streams of information integrate or compete during learning.
Modeling Spatial and Temporal Learning
To tease apart spatial and temporal factors, the researchers employed an encoder-like neural network architecture. The network's task was to predict the identity of a node (an object) and its $(x, y)$ coordinates based on a sequence of "visits" to different locations in a 2D layout.
The architecture consists of:
- Input/Output: $N$ units representing object labels.
- Hidden Layer: 20 units with a tanh transfer function, acting as the network's "internal representation" or "cognitive map."
- Spatial Units: Two additional units representing $(x, y)$ coordinates at the output layer.
The network was trained using a hybrid cost function: cross-entropy for object labels and squared error for spatial coordinates. By adjusting the weights of these errors, the researchers controlled how much the network prioritized spatial versus identity information.
Experimental Conditions and Datasets
The study simulated a scanning process over a synthetic 2D layout of 28 nodes, where the probability of moving to a neighboring node was a function of its distance. The researchers compared three distinct training scenarios:
- Temporal-only (T-only): The network received sequence data but no spatial coordinates.
- Space-Time Consistent: The temporal sequences and the provided $(x, y)$ coordinates matched the same physical layout.
- Space-Time Inconsistent: The temporal sequence was based on the original layout, but the $(x, y)$ coordinates were randomly permuted. This forced the network to handle conflicting information.
Key Findings: The Primacy of Time
The experiments revealed surprising insights into how internal representations are formed:
- Time as a Proxy for Space: Even in the Temporal-only condition, the network developed an internal map. The "distance" between hidden unit activation patterns correlated strongly (approximately 0.6 to 0.7) with the actual physical distance in the layout. The network effectively "discovered" the map simply by observing the sequence of encounters.
- Spatial Information Boosts Generalization: While temporal data alone was powerful, adding consistent spatial coordinates significantly improved the network's ability to generalize. It helped the network "fill in the gaps" for pairs of objects it hadn't seen frequently, creating a more continuous and accurate internal map. The Space-Time Consistent model achieved the highest overall correlation with the actual layout.
- Conflict Resolution: The most striking result emerged from the Space-Time Inconsistent condition. When spatial coordinates conflicted with the temporal sequence, the network's internal representation initially showed some alignment with the spatial data. However, as training progressed, the spatial correlation crashed, and the representation ultimately aligned with the temporal sequence. Time won the tug-of-war.
- Quantitative Proof: A quadratic regression analysis confirmed that the network's internal representation depended much more strongly on temporal distance than on spatial distance.
Limitations and Future Directions
While these findings provide a compelling computational perspective on cognitive mapping, the authors note several limitations. The simulated environment was a small, static layout of 28 nodes, whereas real-world navigation involves vastly more complex and dynamic spaces. Furthermore, while the network uses Hebbian-like associations, backpropagation is a simplification of biological learning mechanisms in areas like the hippocampus.
Finally, the dominance of temporal over spatial information may be task-dependent. Future research could explore alternative loss functions that emphasize active spatial navigation rather than simple identity prediction, potentially shifting the balance back toward spatial coordinates.
Summary updated