A simple model of the co-emergence of grid and place fields

This paper introduces a unified recurrent network model trained on a single sensory-prediction objective that successfully demonstrates the unsupervised co-emergence of grid and place cells, explaining their developmental order and reproducing key experimental phenomena through complementary encoding pressures.

The Gist

Imagine your brain has two special teams of workers that help you navigate the world: the Place Team and the Grid Team.

• The Place Team (found in the hippocampus) acts like a "You Are Here" pin on a map. They light up only when you are in one specific spot, like your living room sofa.
• The Grid Team (found in the entorhinal cortex) acts like graph paper. They light up in a repeating, hexagonal pattern everywhere you go, creating a universal coordinate system to measure distance and direction.

The Big Mystery
For a long time, scientists have been stuck in a "chicken-and-egg" puzzle. The Place Team and the Grid Team are connected and help each other, but how do they both appear during development?

• Does the Grid Team build the map first, and the Place Team just picks a spot?
• Or does the Place Team find a spot first, and the Grid Team builds the map around it?

Most computer models tried to solve this by building one team first and then forcing the other to appear. They didn't quite capture how they grow together naturally.

The New Solution: A Prediction Game
This paper introduces a new computer model that solves the puzzle by treating the brain like a predictive guessing game.

Imagine you are playing a video game where you have to guess what you will see next based on what you just saw and how you moved.

• If you turn left, you expect to see the wall on your left.
• If you move forward, you expect the scenery to shift.

The researchers built a single, unified network of "neurons" (computer nodes) that follows the rule that every neuron is either a "Go" signal (excitatory) or a "Stop" signal (inhibitory)—just like real brains. They trained this network only to be good at guessing the next thing it would see. They didn't tell it, "Hey, be a place cell!" or "Hey, be a grid cell!"

The Magic Result: Co-Emergence
Surprisingly, as the network got better at its guessing game, both types of cells appeared on their own, without any special instructions.

• Some neurons started acting like the Place Team (focusing on specific spots to help reconstruct the current view).
• Others started acting like the Grid Team (focusing on patterns to help predict movement).

It's like a group of people trying to solve a mystery together. Without being told who is the "detective" and who is the "map-maker," they naturally split into those two roles because those are the most efficient ways to solve the puzzle.

Why This Matters
The model didn't just create these cells; it acted exactly like a real animal brain in tricky situations:

1. Hairpin Mazes: When the path loops back on itself, the grid pattern breaks apart, just like it does in real experiments.
2. Removing Walls: When a wall disappears, the grid patterns from different rooms merge, just as they do in reality.
3. Flying Bats: It even recreated the 3D grid patterns seen in bats flying freely.
4. Development: It showed that the "Place" style of thinking tends to appear slightly before the "Grid" style, matching what we see in baby animals growing up.

The Takeaway
The paper suggests that the brain doesn't need two separate blueprints to build these navigation systems. Instead, it just needs one simple goal: predict what happens next.

To do this well, the brain naturally develops two complementary strategies:

1. Reconstruction: "What does this specific spot look like right now?" (Place cells).
2. Prediction: "If I move this way, what will I see next?" (Grid cells).

By playing this single game of "guess the next view," the brain naturally builds both the "You Are Here" pins and the "Graph Paper" map at the same time.

Technical Summary

Technical Summary: A Simple Model of the Co-emergence of Grid and Place Fields

Problem Statement
Spatial navigation in mammals relies on the coordinated activity of grid cells in the medial entorhinal cortex and place cells in the hippocampus. These two cell types are reciprocally connected, creating a developmental "chicken-and-egg" problem: it remains unclear how both cell types arise and reinforce one another during development without pre-existing spatial representations. Existing computational accounts typically suffer from one of two limitations: they derive one cell type strictly from the other, or they model the emergence of a single cell type in isolation using network dynamics. There is a lack of unified models that explain the simultaneous, unsupervised emergence of both grid and place fields.

Methodology
The authors introduce a unified recurrent neural network model designed to address these limitations. The architecture adheres to Dale's Law, ensuring that every neuron is strictly either excitatory or inhibitory, thereby reflecting biological plausibility. The network is trained using a single objective function: predicting the next sensory observation based on masked previous sensory observations and egocentric motion signals.

Crucially, the model does not rely on supervision for either grid or place cells, nor does it depend on pre-existing spatial-cell representations. Instead, the spatial codes are expected to emerge solely from the pressure to minimize prediction error in a sensory-motor context. The model was tested across 1,000 different training configurations to assess the robustness of the co-emergence, with the balance between grid and place coding modulated by varying levels of sensory noise and masking.

Key Contributions

• Unified Co-emergence: This is presented as the first single-objective model where grid and place cells co-emerge without external supervision of either type.
• Biological Constraints: The implementation of Dale's Law distinguishes this model from previous abstract network models.
• Mechanistic Interpretation: The authors propose that the single sensory-prediction objective creates two complementary encoding pressures: (1) correcting errors or reconstructing missing components of sensory observations, and (2) predicting the next sensory state during navigation. These pressures drive the formation of the distinct spatial codes.

Results
The model successfully demonstrates the co-existence of grid and place fields across the majority of training configurations. Without retraining, the network qualitatively reproduces several complex experimental phenomena:

• Grid Fragmentation: Observed in hairpin mazes.
• Grid Merging: Occurring after the removal of walls.
• Lattice Alignment: Across connected rooms.
• 3D Field Ordering: Locally ordered 3D fields similar to those observed in freely flying bats.
• Developmental Sequence: The model reproduces the experimentally observed developmental order where place cells precede grid cells.

The balance between the two spatial codes is shown to be sensitive to the amount of sensory noise and masking applied during training.

Significance
The paper claims to provide a circuit-level account for the co-emergence of grid and place cells, resolving the developmental dependency problem by showing how a single predictive objective can generate both representations simultaneously. Furthermore, the authors assert that their findings generate experimentally testable predictions regarding the behavior of these two kinds of spatial codes, offering a framework to validate the proposed mechanism of sensory-prediction-driven development.