🤖 AI

Geodesic Flow Matching for Denoising High-Dimensional Structured Representations

This paper introduces Geodesic Flow Matching to enforce Riemannian transport dynamics on the toroidal manifold of Spatial Semantic Pointers, thereby overcoming the geometric limitations of Euclidean approaches to significantly improve denoising accuracy and neural efficiency in high-dimensional neurosymbolic SLAM systems.

Original authors: Karim Habashy, Chris Eliasmith

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Karim Habashy, Chris Eliasmith

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Cleaning Up a Messy Map

Imagine you are trying to navigate a city using a magical, high-tech map. This map doesn't use paper; it uses a giant, invisible cloud of data points to represent your location. In the world of "Vector Symbolic Algebras" (a fancy way of saying "math that lets computers think like symbols"), these data points are called Spatial Semantic Pointers (SSPs).

Think of an SSP as a perfectly tuned musical note. To know exactly where you are, the note must have a specific pitch (magnitude) and a specific timing (phase). If the pitch wavers or the timing slips, the computer loses its place on the map.

The Problem: The "Straight Line" Trap

In the real world, things get messy. Just like a musician might cough or a drum might be hit slightly off-beat, these digital maps get "noisy." The data points drift away from their perfect positions.

To fix this, scientists usually try to "clean up" the noise by drawing a line from the messy data back to the correct spot.

The Old Way (Euclidean Flow): Imagine you are on the surface of a perfectly round globe. If you want to get from Point A to Point B, the old method draws a straight line through the center of the Earth.
Why this fails: If you walk through the center of the Earth, you aren't walking on the surface anymore! You lose the "surface" properties. In our musical analogy, this is like trying to fix a note by turning the volume down to zero and then back up. You destroy the delicate timing and pitch structure required to know where you are. The paper shows that for these high-dimensional maps, drawing a straight line through the "middle" ruins the data completely.

The Solution: Walking the Tightrope (Geodesic Flow)

The authors propose a new method called Geodesic Flow Matching.

The Analogy: Instead of drilling through the Earth, imagine you are a tightrope walker. To get from Point A to Point B, you must walk strictly along the curved surface of the globe. This path is called a geodesic.
How it works: The new method forces the "cleaning" process to stay on the surface of the data sphere. It gently guides the messy data back to the correct spot without ever letting it fall into the "void" in the middle. This preserves the delicate timing and pitch (phase and magnitude) that the computer needs to read the map correctly.

The Test: A Spiking Robot Navigator

To prove this works, the researchers built a robot brain using Spiking Neural Networks.

What is that? Think of it as a brain made of tiny, biological-style neurons that fire like little electric sparks. These are very efficient but also very "jittery" and noisy, like a room full of people whispering.
The Challenge: The robot had to navigate a maze (Simultaneous Localization and Mapping, or SLAM) while keeping track of its position. Because the neurons were jittery, the robot's internal map started to drift, like a compass spinning wildly.
The Result: When they used the old "straight line" cleaning method, the robot got lost. But when they used the new "tightrope" (Geodesic) cleaning method:
1. Accuracy: The robot's tracking error dropped by 72%. It stayed on the path much better.
2. Efficiency: The robot could do the same job with 40% fewer neurons. It was like getting a supercomputer's performance out of a simple calculator because the "cleaning" was so efficient.

Summary

The paper argues that when you are cleaning up complex, high-dimensional data that lives on a curved surface (like a sphere), you cannot use standard "straight line" math. You must use "curved path" math. By forcing the cleanup process to stay on the surface of the data sphere, the computer can fix errors without destroying the information, leading to much smarter and more efficient robot navigation.

Technical Summary: Geodesic Flow Matching for Denoising High-Dimensional Structured Representations

1. Problem Statement

Vector Symbolic Algebras (VSAs) enable robust neurosymbolic reasoning by encoding symbolic information into high-dimensional distributed vectors. A critical requirement for these systems is "cleanup"—the ability to map noisy, composed, or partial inputs back to valid, clean states. While well-explored for discrete symbols, cleanup remains an open challenge for continuous representations like Spatial Semantic Pointers (SSPs).

SSPs encode continuous spatial variables into high-dimensional vectors ( $d > 1000$ ) that reside on a highly structured Clifford Hypertorus embedded within the unit hypersphere $S^{d-1}$ . The paper identifies a fundamental geometric failure in applying standard Conditional Flow Matching (CFM) to SSPs:

Euclidean Assumption: Standard CFM assumes a flat Euclidean geometry, defining transport paths as linear interpolants (chords) between a noise sample and a data sample.
Geometric Mismatch: For data constrained to a hypersphere, a straight line connecting two points passes through the interior of the sphere. This causes the vector magnitude to collapse ( $||\phi_t|| < 1$ ) during the transport process.
Consequence: This magnitude collapse destroys the precise phase relationships ( $e^{i\theta_j x}$ ) required to decode the spatial semantics encoded in the SSP. Consequently, Euclidean linear interpolants fail to restore valid SSP states, leading to catastrophic errors in downstream tasks like path integration.

2. Methodology

The authors propose Geodesic Flow Matching (GFM) to resolve the geometric mismatch by adapting Riemannian transport dynamics to strictly restrict the denoising flow to the SSP toroidal manifold.

Core Mechanism

Instead of linear interpolation, GFM defines the probability path along the great-circle arc (geodesic) between the source (noise) and target (clean data) on the unit hypersphere $S^{d-1}$ .

Logarithmic Map ( $\text{Log}_{\phi_0}$ ): Projects the target state $\phi_1$ into the tangent space at the source $\phi_0$ , defining the initial velocity vector.
Exponential Map ( $\text{Exp}_{\phi_0}$ ): Maps a tangent vector back onto the manifold, advancing the state along the geodesic.
Transport Dynamics: The path is defined as $\phi_t = \text{Exp}_{\phi_0}(t \cdot \text{Log}_{\phi_0}(\phi_1))$ .
Training Objective: A neural network $v_\theta(\phi_t, t)$ is trained to regress the instantaneous tangent velocity along this arc. The model minimizes a cosine flow loss to align the predicted velocity with the true geodesic velocity, ensuring updates remain tangent to the hypersphere and preserve the unit-norm constraint by design.

Application Context: Spiking Neural SLAM

The method is validated within a Spiking Neural Network (SNN) implementation of Semantic Simultaneous Localization and Mapping (SLAM).

Challenge: SNNs introduce significant noise (spiking variability, phase drift in recurrent connections) that pushes representations off the valid manifold.
Role of GFM: The GFM model acts as an online stabilizer, intercepting drifting path integration (PI) outputs and projecting them back onto the manifold before they are used for landmark binding. This prevents the accumulation of errors that would otherwise cause trajectory divergence.

3. Key Contributions

Identification of Geometric Failure: The paper demonstrates that standard Euclidean Flow Matching fails for high-dimensional structured representations (SSPs) because linear interpolants "cut through" the manifold's interior, destroying phase and magnitude structure.
Geodesic Flow Matching Framework: The authors adapt flow matching to high-dimensional, toroidal neurosymbolic representations by constraining transport to the Riemannian manifold using logarithmic and exponential maps.
Empirical Validation in SNNs: The approach is tested in a challenging, real-time neurosymbolic benchmark (Semantic SLAM) using spiking neurons, a difficult testbed due to inherent noise.
Efficiency Gains: The method enables high-capacity associative memory in systems where semantics are encoded in vector direction rather than magnitude, bridging the gap between generative transport and valid neurosymbolic cleanup.

4. Experimental Results

The authors compare GFM against Euclidean Flow Matching (CFM), Feedforward Regression, Grid Lookup, and Optimization-based baselines.

Denoising Benchmarks:
- GFM significantly outperforms Euclidean CFM, particularly in high-noise regimes where linear interpolants fail to recover the ground truth.
- In high-noise conditions (Signal Strength < 0.15), GFM outperforms non-neural baselines (Grid/Optimization) which fail to "snap" to correct prototypes once the signal drops below the noise floor.
- Qualitative analysis shows GFM reconstructs sharp, unimodal peaks at the correct spatial coordinate, whereas Euclidean models produce spatially displaced "drift" and feedforward models produce diffuse, averaged representations.
SLAM Performance (Spiking Neural System):
- Error Reduction: In a resource-constrained regime ( $N=1500$ neurons), GFM reduced the average path tracking error by 72% compared to a Grid baseline (0.076m vs. 0.249m).
- Neural Efficiency: A system with 1,500 neurons augmented with GFM achieved tracking accuracy comparable to a baseline system using 2,500 neurons. This represents a 40% increase in neural efficiency (or a 40% reduction in required resources) for stable path integration.
- Stability: Unlike Grid baselines that introduce discontinuous state jumps (causing trajectory divergence in dynamic integration), GFM provides a continuous correction field that maintains stable long-term integration.

5. Significance and Claims

The paper claims that the intrinsic geometry of high-dimensional representations is critical for robust cleanup. By shifting the paradigm from discrete search in the embedded domain to continuous generative transport directly in the embedding space, the authors provide a robust framework for hyperspherical embeddings.

The significance lies in:

Geometric Consistency: Ensuring transport remains consistent with the manifold prevents the destruction of phase information, which is essential for accurate decoding in VSAs.
Neuromorphic Efficiency: The method enables efficient neuromorphic computing by stabilizing noisy spiking representations, allowing smaller networks to perform tasks that previously required larger resources.
Generalizability: While focused on SSPs, the approach applies to other architectures utilizing hyperspherical embeddings, such as Hyperspherical Prototype Networks and Hyperspherical Variational Autoencoders, where vector direction encodes semantics.

The authors note that while the current architecture prioritizes validating the geometric formulation over compute minimization, the framework is designed to be convertible to fully spiking networks for deployment on physical neuromorphic hardware.