A Vector Navigation and Inference Architecture can Construct Universal Cognitive Maps for Abstract Reasoning

This paper proposes a neurally plausible model demonstrating that spatial navigation architectures, specifically those involving grid cells, can construct universal cognitive maps to support domain-general abstract reasoning tasks such as analogy making and perspective taking.

The Gist

Imagine your brain has a built-in GPS. For decades, scientists thought this GPS was only used for one thing: helping you find your way home, navigate a city, or remember where you left your keys. This system relies on special "grid cells" that light up like a hexagonal honeycomb whenever you move through space.

But this new paper suggests something much more exciting: Your brain's GPS is actually a universal mapmaker for everything, not just physical space.

Here is the story of how the brain might turn abstract ideas into a map, explained simply.

1. The Problem: How do we map "ideas"?

We can easily map physical space. If you walk 10 steps North, you are 10 steps away from where you started. But how do you map something abstract, like emotions?

• Is a "happy" feeling close to a "calm" feeling?
• Is a "scary" feeling far away from a "boring" one?

Scientists have found that the brain's grid cells also light up when we think about non-physical things, like the "valence" (good vs. bad) and "arousal" (calm vs. excited) of an image, or even the shape of a bird's neck and legs. But how does the brain build a map for these invisible concepts?

2. The Solution: The "Universal Cognitive Map"

The author, Andrej Bicanski, proposes a clever trick. He suggests the brain uses the exact same machinery it uses for walking around a room to "walk" around a concept.

Think of it like this:

• The Map: Imagine a giant, invisible sheet of graph paper (the grid) in your brain.
• The Anchor: You take your first idea (e.g., "a very happy dog") and pin it to a random spot on the graph paper.
• The Vector (The Arrow): You look at a second idea (e.g., "a slightly less happy cat"). You calculate the "distance" between them. If they are very similar, the arrow between them is short. If they are very different, the arrow is long.
• The Magic Step: The brain uses a special "Positional Inference Network" (let's call it the Map-Maker) to say: "Okay, I know where the dog is. I know the arrow to the cat is short and points right. Therefore, the cat must be pinned right here on the graph paper."

By repeating this over and over, the brain builds a complete, organized map of the concept, even if the concepts are abstract.

3. The "Bird" and "Mood" Experiments

The paper tested this idea with two examples:

• The Mood Map (OASIS Dataset): The researchers took thousands of images rated for how "happy/sad" and "calm/exciting" they were. They fed these into the model. The model successfully pinned every image to a spot on the grid. When they looked at the result, the images formed a perfect hexagonal honeycomb pattern, just like a real map of a city!
• The Bird Map: They created a world of "stretchy birds" with different neck and leg lengths. The model mapped these birds based on their body shapes. Again, the birds organized themselves perfectly on the grid, with long-necked birds in one corner and short-legged birds in another.

4. Why is this a big deal? (The "Reasoning" Superpower)

Once the map is built, the brain can do amazing things without needing a new brain part. It can use the map to reason.

• Analogy Making: Imagine you know the relationship between a Cat and a Kitten (it's a "smaller version" vector). If you want to know what a Dog is to, you just apply that same "smaller version" arrow to the Dog on your map. Poof! You find the Puppy. The brain can solve "A is to B as C is to...?" just by drawing arrows on its internal map.
• Perspective Taking: You can imagine yourself standing on the map. If "Success" is to your left and "Failure" is to your right, the brain can tell you, "I am currently feeling like I'm on the 'Success' side."
• Finding Shortcuts: Just like a GPS finds a shortcut through a park you've never walked through, this system can figure out the relationship between two ideas you've never directly compared before, just by calculating the vector between them.

5. The "Noise" Factor: What if the brain is messy?

Real life is messy. Sometimes our memories are fuzzy, or our feelings are hard to pin down. The paper shows that this map system is fault-tolerant.

• If one calculation is slightly wrong (noise), the brain doesn't crash.
• Instead, it asks for more "anchors." It checks the position of a new idea against three or four existing ideas on the map instead of just one.
• If three out of four say "It goes here," the brain ignores the one outlier. It's like asking a group of friends for directions; if three say "Left" and one says "Right," you go Left.

The Big Picture

This paper suggests that navigation is the foundation of thought.

We don't need a separate "logic machine" in our brains to do abstract reasoning. We just repurpose our ancient, evolutionary "GPS." By turning concepts into locations on a map, and relationships into arrows (vectors), our brains can navigate the world of ideas just as easily as we navigate the world of streets.

In short: Your brain doesn't just know where you are in the city; it knows where you are in the world of emotions, logic, and imagination, using the same old map.

Technical Summary

1. Problem Statement

The concept of the "cognitive map" has traditionally been rooted in spatial navigation, supported by the discovery of grid cells in the entorhinal cortex and place cells in the hippocampus. However, recent evidence suggests these systems also support non-spatial, abstract cognition (e.g., social hierarchies, conceptual spaces, affective states).

The core problem addressed in this paper is how to construct a universal cognitive map (UCM) for arbitrary, abstract stimulus dimensions using the neural machinery of spatial navigation. Specifically:

• How can the brain map discontinuous, abstract stimulus spaces (like "valence" and "arousal" or "leg length" and "neck length") onto the periodic, toroidal neural manifolds of grid cells?
• How can the system infer the position of a new abstract stimulus based on its similarity to known stimuli without continuous physical movement?
• How can this architecture support higher-order reasoning (analogy, perspective taking) using the same mechanisms used for map construction?

2. Methodology

The author proposes a computational model comprising three core components that interact to build and utilize UCMs:

A. Core Architectural Components

1. Grid Cell Network: A standard multi-scale grid cell system (6 modules with varying scales and orientations) representing a toroidal manifold. Locations are encoded as Population Vectors (PVs) across these modules.
2. Vector Navigation Architecture (VNA): A mechanism (based on distance cells) that computes the displacement vector between two known locations (A and B) on the grid map.
3. Positional Inference Network (PIN): A novel component that performs the inverse operation of the VNA. Given a starting location (A) and a displacement vector (A→B), the PIN infers the target location's grid cell PV (B). This allows the system to "jump" to a new location on the map based on abstract similarity metrics.

B. The Mapping Algorithm

The construction of a UCM follows an iterative triangulation process:

1. Anchoring: The first stimulus is anchored to an arbitrary position on the grid map.
2. Vector Calculation: The difference in magnitude between the first stimulus and a second stimulus (along relevant dimensions) is treated as a displacement vector.
3. Inference: The PIN uses the anchor's PV and the displacement vector to infer the PV for the second stimulus.
4. Triangulation: As more stimuli are introduced, the system uses multiple existing anchors to infer the position of a new stimulus. If multiple inferences converge (overlap), the stimulus is anchored; if they diverge, the system can reject the placement or average the coordinates to handle noise.
5. Plastic Gain: The model assumes a plastic "velocity gain" that scales the abstract stimulus dimensions to fit the grid cell rate maps, allowing the map to adapt to different ranges of data.

C. Datasets and Validation

The model was tested on two datasets:

1. OASIS (Open Affective Standardized Image Set): A dataset of images rated on Valence and Arousal. This tests the mapping of affective states.
2. Synthetic "Bird-Space": A procedurally generated dataset varying Leg Length and Neck Length. This tests the mapping of physical feature dimensions.

D. Noise Handling

The model incorporates a noise-compensation mechanism. When inferring a new position, the system can use $N$ different anchors. If the inferred positions vary due to noise, the system takes the mean/median of the valid inferences within a tolerance threshold. If no consensus is reached, the stimulus is temporarily rejected, requiring more trials (anchors) to resolve.

3. Key Contributions

• Universal Cognitive Map Construction: Demonstrates that spatial navigation architectures (Grid Cells + VNA) can be repurposed to construct metric maps for non-spatial, abstract dimensions.
• Positional Inference Network (PIN): Introduces a neurally plausible mechanism for inferring target positions from vectors, a capability previously unexplored in grid cell literature.
• Discontinuous Mapping: Shows that UCMs can be built from discontinuous stimulus sets (jumping between abstract concepts) without requiring smooth, continuous trajectories, unlike previous "mental navigation" models.
• Unified Reasoning Framework: Proposes that the same mechanisms used to build the map (VNA and PIN) are sufficient to perform complex reasoning tasks, eliminating the need for separate cognitive modules.

4. Results

• Successful Map Construction: The model successfully mapped both the OASIS (affective) and Bird-Space (feature) datasets onto the grid cell manifolds.
• Gridness Recovery: "Recovered" grid cells (visualized by correlating inferred positions with ground truth) exhibited strong hexagonal firing patterns and high "gridness" scores, particularly in smaller scale modules.
• Noise Tolerance: The system demonstrated robustness. While high noise degraded the resolution of small-scale grid modules, increasing the number of triangulation anchors (retries) restored gridness and allowed the system to anchor stimuli correctly.
• Reasoning Capabilities: The model successfully simulated:
  • Analogical Reasoning: Solving "A is to B as C is to D" by computing vector A→B and applying it to C to find D.
  • Egocentric Perspective Taking: Calculating the relative bearing of a target from a "self" position using a heading vector.
  • Subspace Construction: Identifying subsets of stimuli that lie along a specific linear trajectory (e.g., all birds where leg length equals neck length) by scaling vectors.

5. Significance and Implications

• Domain-General Cognition: The paper provides a strong theoretical argument that the hippocampal-entorhinal system is a domain-general substrate. It suggests that the neural machinery evolved for spatial navigation is the fundamental engine for abstract reasoning and relational inference.
• Behavioral Predictions: The model predicts that difficult abstract tasks (high noise) will require more trials to learn the underlying structure, mirroring findings in human behavioral experiments. It also predicts that spatial navigation deficits should correlate with deficits in abstract reasoning.
• Neural Plausibility: By relying on known cell types (grid cells, distance cells, head direction cells) and plausible synaptic mechanisms (gating), the model bridges the gap between abstract computational theories and biological reality.
• Reinterpretation of Grid Cells: It suggests that grid cells do not merely represent physical space but represent a metric for any systematically varying dimension, with the "velocity" being the rate of change in abstract feature space.

In conclusion, Bicanski's work offers a unified architecture where navigation is inference, and inference is navigation, suggesting that the brain uses a single, flexible vector-based system to map and reason about the world, whether that world is physical or conceptual.