From movement to cognitive maps: recurrent neural networks reveal how locomotor development shapes hippocampal spatial coding

1. Problem Statement

The hippocampus contains spatially tuned neurons (place cells, head direction cells, grid cells) that form a "cognitive map" essential for navigation. While the timeline of their postnatal development in rats is well-characterized (e.g., head direction cells emerge first, followed by place cells, then grid cells), the mechanistic drivers behind this sequential emergence remain unclear.

Existing theories suggest that sensory experience shapes these representations, but it is unknown whether:

1. The specific statistics of locomotor development (how movement patterns change from crawling to walking to running) are the causal factor.
2. Simply increasing the rate of sensory change (e.g., faster visual updates) is sufficient to drive this maturation.
3. The transition from egocentric (self-centered) to allocentric (world-centered) coding is driven by the complexity of movement trajectories.

2. Methodology

The authors employed a computational neuroscience approach, combining the analysis of experimental behavioral data with a Recurrent Neural Network (RNN) model trained on a predictive task.

A. Computational Analysis of Locomotor Development

• Data Aggregation: The authors aggregated open-field exploration data from rats aged Postnatal Day (P) 11 to P25 and adults from multiple previous studies.
• Clustering: They computed probability density functions (PDFs) for speed and rotational speed, along with spatial transition matrices. Using Jensen-Shannon distances and t-SNE visualization, they identified three distinct locomotor stages based on movement statistics:
  1. Crawl: Median age ~P13.5 (uncoordinated, slow).
  2. Walk: Median age ~P16 (emerging quadrupedal gait).
  3. Run: Median age ~P20 (faster, more stable).
  4. Adult: 3–6 months.
• Trajectory Simulation: Using the RatInABox toolbox, they generated synthetic trajectories matching the statistical properties (speed, rotational velocity, thigmotaxis) of each developmental stage.

B. The RNN Model (Artificial Hippocampus)

• Architecture: A single-layer RNN with 500 hidden units.
• Input: At each timestep $t$, the network received a concatenation of:
  • Visual Input ($Y_t$): Low-resolution (32x16) panoramic frames (240° FOV) simulating rat vision.
  • Vestibular Input ($v_t, \omega_t$): Velocity vector and rotational speed.
• Objective: The network was trained to predict the next visual frame ($\tilde{Y}_{t+1}$) based on current inputs and hidden state. This forces the network to build an internal model of the environment to anticipate future sensory inputs.
• Training Protocol:
  • Sequential Training: Networks were trained sequentially across developmental stages (Crawl $\to$ Walk $\to$ Run $\to$ Adult). Weights from the previous stage were used to initialize the next (simulating continuous learning).
  • Grid Cell Variant: For the "Adult" stage, a variant was tested where the hidden state was periodically initialized with a linear projection of 25 pre-defined grid cells to test if grid cell input is necessary for mature place cell tuning.

C. Analysis and Validation

• Spatial Tuning Metrics: Hidden units were analyzed for:
  • Place Cells: Spatial Information ($SI_r$) on rate maps.
  • Head Direction (HD) Cells: Resultant Vector Length (RVL) and Spatial Information ($SI_d$) on polar maps.
  • Conjunctive Cells: Units exhibiting both spatial and directional tuning.
• Comparison: Model outputs were compared against real hippocampal CA1 recordings from developing rats.
• Control Experiments:
  • Rate of Change Control: Training on "Crawl" trajectories but with extended inter-frame gaps (1–3s) to mimic the predictive difficulty of adult movement without changing movement statistics.
  • Reverse Training: Training from Adult $\to$ Crawl.
  • Lesion Studies: Silencing spatially tuned units to measure impact on task performance.

3. Key Contributions

1. Mechanistic Link: The paper establishes a causal link between locomotor statistics and the ontogeny of spatial coding. It demonstrates that the specific progression of movement complexity (crawl $\to$ walk $\to$ run) is required to drive the sequential emergence of spatial representations.
2. Novel Prediction (Conjunctive Coding): The model predicts that the emergence of directional selectivity in the hippocampus occurs primarily through conjunctive place-direction cells (neurons that encode both location and heading) rather than purely directional cells.
3. Disproving "Rate of Change" Hypothesis: The study proves that simply increasing the temporal gap between sensory inputs (making the prediction task harder) is insufficient to generate adult-like spatial representations. The specific kinematics of movement are critical.
4. Grid Cell Dependency: The model suggests that while early spatial tuning emerges from movement statistics alone, mature, stable place cell tuning in adults may depend on the integration of grid cell inputs.

4. Key Results

• Recapitulation of Developmental Timeline:
  • RNNs trained on "Crawl" data showed minimal spatial tuning.
  • "Walk" stage training produced the emergence of Head Direction (HD) cells.
  • "Run" stage training led to the emergence of Place Cells.
  • The model's progression of spatial information ($SI_r$) and directional metrics ($SI_d$, RVL) closely matched the experimental data from real rats (P14–P25).
• Conjunctive Cell Emergence:
  • The model predicted a significant increase in conjunctive place-HD cells during development.
  • Experimental Validation: The authors re-analyzed existing CA1 recordings and confirmed a significant developmental increase in conjunctive place-direction cells, a finding previously unreported in developmental literature.
• Failure of Rate-of-Change Controls:
  • Models trained on "Crawl" trajectories with artificially extended time gaps (to simulate adult predictive demands) failed to develop robust spatial representations. They showed higher decoding errors and lower spatial information compared to models trained on naturalistic developmental trajectories.
  • This confirms that the structure of movement, not just the frequency of sensory updates, drives cognitive map formation.
• Lesion and Robustness:
  • Silencing spatially tuned units in the model caused a significant increase in validation loss, confirming their functional necessity for the predictive task.
  • The results were robust to variations in network size, input dimensions, and training duration.

5. Significance

• Embodied Cognition: The work provides strong evidence for embodied cognition in neural development, showing that the physical act of moving (locomotor statistics) is not just a byproduct but a fundamental driver of how the brain constructs spatial maps.
• Developmental Disorders: By linking motor development to cognitive map formation, the findings offer new insights into neurodevelopmental disorders where motor and spatial deficits co-occur (e.g., autism or specific learning disabilities).
• AI and Robotics: The study suggests that for artificial agents to develop robust, allocentric spatial representations, they must experience progressively complex motor behaviors rather than just being fed high-frequency sensory data.
• Theoretical Advancement: It challenges the view that spatial coding is purely a result of sensory input statistics, highlighting the critical role of motor output statistics in shaping the internal dynamics of recurrent networks.

In summary, the paper demonstrates that the "cognitive map" is not pre-wired but is sculpted by the specific, evolving statistics of an animal's movement through the world, with recurrent neural networks serving as a powerful tool to uncover these developmental mechanisms.