Research map / Continual learning / Online and replay-based
Rapid Learning without Catastrophic Forgetting in the Morris Water Maze
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the sequential Morris Water Maze (sWM) task and a neuro-inspired model (vHSN) built on Vector-HaSH grid/place codes, a fixed random CNN displacement network, and attention memory, which avoids catastrophic forgetting across environments better than continual learning baselines.
Problem
Continual learning and rapid adaptation in navigation: standard ANNs forget previously learned environments when trained sequentially and do not transfer quickly to new goal locations or environments.
Approach
Combine a Vector-HaSH content-addressable entorhinal-hippocampal memory (grid codes shared across environments) with a randomly initialized fixed spatially invariant CNN displacement network and an attention key-value memory, so new environments need no policy retraining (zero-shot transfer) and past knowledge is retained.
Key results
- vHSN achieves 99.2% average success rate across five environments after sequential training, versus 19.5% for fine-tuning and 23.2% for EWC.
- Replay-based and architecture-based baselines (ER, DER, DER++, A-GEM) reach between 2.05% and 16.92% average success rate.
- Attention alone on fixed goal reaches 100% but fails on new goal locations (about 1.6-2% success), while the full combination reaches 99.5% on new goal locations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use vHSN instead of EWC | sWM · Average across environments after all training | success rate (%): 23.2 -> 99.2 | Helped |
| Add attention to Vector-HaSH and displacement network (ablation) | sWM · All environments after training | success rate (%): 5.4 -> 99.2 | Helped |
| Add Vector-HaSH to displacement network and attention (ablation) | sWM · All environments after training | success rate (%): 0.9 -> 99.2 | Helped |
| Add the displacement network to Vector-HaSH and attention (ablation) | sWM · All environments after training | success rate (%): 99.7 -> 99.2 | No clear effect |
| Add the spatially invariant CNN displacement network to Vector-HaSH and attention (ablation) | sWM · New goal location after training | success rate (%): 1 -> 99.5 | Helped |
| Add attention to Vector-HaSH and displacement network (ablation) | sWM · New goal location after training | success rate (%): 1.8 -> 99.5 | Helped |
| Replace the CNN displacement network with fully connected layers (ablation) | Vector-HaSH + attention · New goal location after training | success rate (%): 99.5 -> 0.1 | Hurt |
Surprising findings
- Attention module alone achieves perfect success on a fixed goal location but collapses under goal changes or new environments, showing memorization is not transfer.
- Standard continual learning baselines including EWC and replay perform poorly, with EWC attributed to similar observations mapping to different actions across environments.
Limitations
- Evaluated only on a simplified simulated square maze with four cardinal actions and noise-added wall markings.
- Vector-HaSH model is trained only on a single environment, and the approach's scalability to more complex dynamic environments is untested.
- Ablated models were trained and evaluated differently (single-environment training for the full framework), complicating direct comparison.
Experimental setup
Datasets: sequential Morris Water Maze (sWM) simulated environments
Benchmarks: sequential Morris Water Maze (sWM)
Metrics: average success rate (%)
Baselines: Fine-tuning EWC Replay buffer DER DER++ A-GEM ER
Builds on: Vector-HaSH Elastic Weight Consolidation Dark Experience Replay
Closest papers by meaning
- Thalamus: a brain-inspired algorithm for biologically-plausible continual learning and disentangled representations (ICLR 2023 · similarity 0.88)
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.87)
- Predictive Attractor Models (NeurIPS 2024 · similarity 0.86)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.86)
- Memory-Integrated Reconfigurable Adapters: A Unified Framework for Settings with Multiple Tasks (NeurIPS 2025 · similarity 0.86)
- Flexible task abstractions emerge in linear networks with fast and bounded units (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.