New AI model reveals how grid and place cells co-emerge for navigation
A single neural network replicates decades of spatial mapping experiments without labeled data.
The authors built a unified recurrent network where grid and place cells co-emerge from a single sensory-prediction objective. Trained to predict the next sensory observation from masked previous sensory observations and egocentric motion, the model reproduces grid fragmentation in hairpin mazes, grid merging after wall removal, and locally ordered 3D fields observed in freely flying bats across 1,000 different training configurations — without pre-existing spatial representations or supervision of either cell type.
- First unified model where grid and place cells co-emerge without any supervision of either type, using only a sensory-prediction objective.
- Reproduces 6 distinct experimental phenomena (grid fragmentation, merging, alignment, 3D bat fields, developmental order) across 1,000 training configurations without retraining.
- Model instantiates Dale's Law (excitatory/inhibitory neurons) and uses a single training objective — predictive coding of sensory inputs and egocentric motion.
Why It Matters
This model provides a unified computational framework for spatial navigation, linking neural circuit dynamics to self-supervised AI learning.