Hippocampal Entorhinal World Model

---
name: hippocampal-entorhinal-world-model
description: "Brain-inspired hierarchical world model using hippocampal-entorhinal (HPC-MEC) circuit for structure abstraction and generalization. Inverse model for structural extraction, HPC-MEC coupling for dissociating relational structures from episodic scenes. Based on Zhang et al. (arXiv: 2605.15733). Use when designing brain-inspired world models, building self-supervised learning systems with structural generalization, modeling hippocampal-entorhinal computation, or developing abstraction mechanisms for spatial/conceptual representations."
---

# Hippocampal-Entorhinal World Model for Structure Abstraction

> Brain-inspired hierarchical model that concurrently extracts abstract structures from continuous, high-dimensional dynamics using hippocampal-entorhinal (HPC-MEC) circuit architecture, enabling structural abstraction and generalization through velocity-driven path integration.

## Metadata
- **Source**: arXiv:2605.15733
- **Authors**: Tianqiu Zhang, Muyang Lyu, Xiao Liu, Si Wu
- **Published**: 2026-05-15
- **Categories**: cs.NE, cs.AI, cs.CV

## Core Methodology

### Key Innovation
While the hippocampal-entorhinal (HPC-MEC) circuit is known to represent both spatial and conceptual spaces, the mechanisms for **concurrently extracting abstract structures from continuous, high-dimensional dynamics** remain poorly understood. This work proposes a brain-inspired hierarchical world model that:

1. **Simultaneously infers latent transitions and constructs a predictive visual world model**
2. **Dissociates relational structures (MEC) from integrated episodic scenes (HPC)**
3. **Achieves structural generalization via velocity-driven path integration**

### Technical Framework

#### 1. Inverse Model for Structural Extraction
- Learns to infer latent state transitions from observed dynamics
- Acts as the structural extraction mechanism analogous to MEC grid cell computation
- Maps high-dimensional observations to abstract relational representations
- Self-supervised: no external labels required

#### 2. HPC-MEC Coupling Model
**Medial Entorhinal Cortex (MEC) Pathway:**
- Extracts and maintains **relational structures**
- Analogous to grid cell encoding of spatial/conceptual relationships
- Provides a structured latent space for transitions
- Velocity-driven path integration for structural generalization

**Hippocampal (HPC) Pathway:**
- Integrates relational structures with sensory information into **episodic scenes**
- Analogous to place cell formation from grid cell inputs
- Combines structural priors with perceptual content
- Enables predictive visual world modeling

#### 3. Predictive Visual World Model
- Uses the coupled HPC-MEC representation for next-state prediction
- Velocity signals drive path integration across abstract spaces
- Enables **structural reuse** across diverse contexts
- Supports transfer of learned structures to novel environments

#### 4. Velocity-Driven Path Integration
- Core mechanism for structural generalization
- Abstract velocity signals in latent space drive state transitions
- Allows model to "navigate" conceptual spaces analogously to spatial navigation
- Enables prediction and structural reuse without retraining

### Benchmark: Primitive Transformation Dynamics
- Demonstrates capacity for structural abstraction on controlled benchmark
- Tests ability to extract invariant structures from transforming inputs
- Shows robust prediction across diverse contexts

## Implementation Guide

### Architecture Overview
```
Input (high-dim dynamics)
         |
         v
    [Inverse Model]  <-- Structural extraction
         |
         v
    [HPC-MEC Coupling]
     /            \
    v              v
[MEC: Relations]  [HPC: Episodic Scenes]
    \            /
     v          v
[Predictive World Model]
         |
         v
    Next-State Prediction
```

### Key Components

#### Inverse Model
```python
class InverseModel(nn.Module):
    """Infers latent transitions from observed dynamics."""
    def forward(self, x_t, x_tp1):
        # Extract structural transition from state pair
        latent_transition = self.encoder(torch.cat([x_t, x_tp1], dim=-1))
        return latent_transition
```

#### HPC-MEC Coupling
```python
class HPOMECCoupling(nn.Module):
    """Dissociates relational structures from episodic scenes."""
    def __init__(self, latent_dim, scene_dim):
        super().__init__()
        self.mec_path = nn.Sequential(...)  # Relational structure
        self.hpc_path = nn.Sequential(...)  # Episodic integration
        self.velocity_encoder = nn.Sequential(...)  # Path integration
        
    def forward(self, latent_transition, sensory_input, velocity):
        # MEC: extract relational structure
        relational = self.mec_path(latent_transition)
        
        # HPC: integrate with sensory for episodic scene
        episodic = self.hpc_path(relational, sensory_input)
        
        # Velocity-driven path integration
        integrated = self.velocity_encoder(velocity, relational)
        
        return relational, episodic, integrated
```

### Prerequisites
- PyTorch for model implementation
- Understanding of hippocampal-entorhinal circuit biology
- Experience with self-supervised learning and world models

## Applications
- **Brain-inspired world models**: Self-supervised learning with structural priors
- **Spatial/conceptual representation learning**: Grid cell-inspired encoding
- **Structural generalization**: Transfer of abstract knowledge across contexts
- **Robotics navigation**: Velocity-driven path integration in latent spaces
- **Cognitive AI**: Models of hippocampal function for episodic memory

## Pitfalls
- Requires carefully designed velocity signals for effective path integration
- Structural abstraction quality depends on inverse model capacity
- HPC-MEC dissociation may need regularization to prevent collapse
- Validation on complex, real-world dynamics remains challenging
- Limited to environments with exploitable structural regularities

## Related Skills
- hippocampal-replay-credit-assignment
- vacoal-algebro-deterministic-memory
- brain-inspired-memory-ai-agents
- neat-nc-navigation-cells
- brain-inspired-memory-agents