VectorHaSH

A biologically-plausible vector hashing framework for associative memory and bio-inspired SLAM, developed as part of the RatSLAM 2.0 COMP 400 honors project at McGill University.

Overview

VectorHaSH provides a neocortical-hippocampal-entorhinal memory scaffold that stores sensory observations in a distributed, denoising-friendly representation. The framework is built around:

• vectorhash.py – the core VectorHaSH class (store_memory, recall, denoising, and dream/fix modes).
• clean_scaffold.py – GridHippocampalScaffold modeling entorhinal grid cells over a toroidal attractor network.
• hippocampal_sensory_layers.py – HippocampalSensoryLayer for Hebbian / analytic / iterative pseudoinverse learning between the hippocampus and sensory inputs.
• shifts.py – velocity-driven continuous attractor dynamics for path integration.
• graph_utils.py / grid_module.py – topological graph building and grid-module utilities used for SLAM loop-closure and pose tracking.

Experiments (capacity scaling, learning-rule benchmarks, and SLAM evaluations in AnimalAI) are in the notebooks and experiments/, kidnapping_tests/, and results/ directories.

Visuals

RatSLAM 2.0 pose tracking in AnimalAI — a test path with the agent's true vs. predicted camera view, plus the estimated x, y, and heading distributions tracked by the hippocampal grid code:

Grid-module dynamics — the entorhinal grid-cell continuous attractor network (CAN) sustains a localized activity bump that drifts with velocity input (path integration). Left: a bump circulating under rotational velocity input. Right: bump stabilization/repair via the Burak & Fiete healer.

circular velocity drive	stabilization by healer

Shift dynamics — how new sensory evidence moves the grid-cell bump: additive shift vs. multiplicative shift of the one-hot grid representation.

additive shift	multiplicative shift

Write-up

Full write-up, abstract, and presentation slides: Research & Publications – RatSLAM 2.0 (VectorHaSH & Bio-Plausible SLAM)

Dependencies

Python + PyTorch, with scipy, numpy, and matplotlib. See environment-aai-linux.yml for a full conda environment.

Citation

@techreport{huang2025ratslam2,
  title={COMP 400 Final Report: RatSlam 2.0 (VectorHaSH \& Bio-Plausible SLAM)},
  author={Huang, Ezra and Lamontagne-Kratz, Jonathan and Lefevre, Olivier and Premont-Schwarz, Isabeau},
  institution={McGill University School of Computer Science},
  type={COMP 400 Final Report},
  month={May},
  year={2025},
  url={https://github.com/epicgamer17/vectorhash}
}

License

MIT

Note: the kidnapping_tests/, kidnapping_tests_identity_smoothing/, results/, and hypers_searching/ directories contain generated experiment outputs (.pkl checkpoints, .zip archives) and are gitignored. They are recreated by running the notebooks.