VSA-OGM
Compressing a spatial map into high-dimensional memory, then querying it to recover occupancy.
What
How can neural representations encode and recall spatial information? VSA-OGM explores this question by encoding spatial observations into high-dimensional vector memories, then querying them to reconstruct a map.
Why
A useful spatial memory should retain what has been observed and answer questions about a location without storing every observation separately.
How
Encode points as vectors, combine them in memory, then query locations to recover an occupancy grid map (OGM).
What’s next
How should a memory change when new information arrives? SRMU: relevance-gated memory updates →
From points to memory to a map
The central idea is to compress a map into a vector memory and reconstruct spatial information by querying it. The published implementation uses separate occupied/empty memories and spatial tiles to manage capacity.
1. Encode points
Each coordinate becomes a high-dimensional vector that represents its position.
2. Store in memory
Combine point vectors into occupied and empty memories for each tile.
3. Query the map
Compare a query vector with the memories to estimate occupancy at that location.
A closer look
Results: Intel Map
Intel Map tests the approach on recorded indoor spatial observations. At 0.1 m grid resolution, the paper reports:
| Method | AUC ↑ | CPU time per point cloud ↓ | Model size |
|---|---|---|---|
| VSA-OGM | 0.95 | 0.51 s | 16.3 MB |
| BHM, full covariance | 0.96 | 22.35 s | 6,400 MB |
| BHM, diagonal covariance | 0.96 | 8.11 s | 0.04 MB |
This setting trades a small AUC difference for lower processing time than either BHM variant. The memory benefit is relative to full covariance; the diagonal model is smaller. AUC measures how well occupied and empty locations are distinguished; higher is better.
These are reported measurements, with point-cloud processing and map generation included. The workstation used an Intel Xeon W-2295 CPU, 128 GB RAM, and an RTX A5000 GPU; the table above uses CPU timings. See Table 6 in the published paper.
For context, on the synthetic ToySim dataset, VSA-OGM scored 0.93 AUC versus 0.96/0.97 for full/diagonal BHM, with negative log-likelihood of 0.33 versus 0.44/0.38. Lower negative log-likelihood is better. See Table 4.
Capacity and limitations
Uneven spatial sampling can degrade the map. Memories also saturate as observations accumulate, requiring larger vectors or more tiles. The goal is to represent a map in a compact memory, but how much detail it can retain depends on the memory’s capacity. See the Intel Map discussion.
My contribution
I developed the idea, designed and implemented the system, ran the experiments, analyzed the results, and wrote the paper, with the advice and guidance of my coauthors Andrew Capodieci, David Gorsich, and Maryam Parsa.
The published contribution statement records their roles in the idea, analysis, review, writing, and project support.
Related papers
SRMU: Relevance-Gated Updates for Streaming Hyperdimensional Memories
Shay Snyder, Andrew Capodieci, David Gorsich, and Maryam Parsa · International Conference on Neuromorphic Systems (ICONS) · 2026 · Conference
SRMU studies how to update hyperdimensional memories as new information arrives, using relevance to decide what to retain.
Paper details and citation →Paper and code
Brain Inspired Probabilistic Occupancy Grid Mapping with Vector Symbolic Architectures. npj Unconventional Computing, 2026.
Read the paper on this page
Published version by Shay Snyder, Andrew Capodieci, David Gorsich, and Maryam Parsa. Reproduced unchanged under CC BY-NC-ND 4.0.
Cite this work
@article{snyder2026brain,
title = {Brain Inspired Probabilistic Occupancy Grid Mapping with Vector Symbolic Architectures},
author = {Snyder, Shay and Capodieci, Andrew and Gorsich, David and Parsa, Maryam},
journal = {npj Unconventional Computing},
volume = {3},
pages = {13},
year = {2026},
doi = {10.1038/s44335-026-00052-w}
}