Research map / Architectures and efficiency / Attention and transformers
Hippoformer: Integrating Hippocampus-inspired Spatial Memory with Transformers
Tiantian Li, Xingxing Cao, Yifei Wang, Xiaojiao Yang, Xiaolong Zou, Bo Hong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace TEM with mm-TEM (hierarchical MLP relational memory) | TEM · 2D grid prediction | test accuracy / training efficiency: about 60% after 20,000 steps -> nearly 90% within 5,000 steps | Improved |
| add auxiliary relational loss L_g2g (ablation) | mm-TEM · 2D grid prediction | generalization accuracy | Improved |
| add auxiliary relational loss L_s2g (ablation) | mm-TEM · 2D grid prediction | generalization accuracy | Improved |
| add all auxiliary relational memory losses (ablation) | mm-TEM · 2D grid prediction | generalization accuracy | Improved |
| combine one-layer Transformer with mm-TEM (Hippoformer) instead of mm-TEM alone | mm-TEM · 3D environment 1-step error | error (1e-3): 5.10 -> 1.27 | Improved |
| combine mm-TEM with Transformer instead of Transformer alone | Transformer · 3D environment multi-step imagination | error (1e-3): 36.13 -> 9.71 | Improved |
| combine mm-TEM with Transformer instead of Titans | Titans · 3D environment multi-step imagination | error (1e-3): 33.42 -> 9.71 | Improved |
| combine Transformer with mm-TEM versus Transformer alone for 1-step | Transformer · 3D environment 1-step error | error (1e-3): 1.29 -> 1.27 | No clear effect |
| increase memory update interval mb (ablation) | mm-TEM · 2D grid prediction | grid scale: coarser grids at larger mb; single-step accuracy reduced | Mixed |
| add mm-TEM path integration with higher grid score | mm-TEM · 2D grid multi-step accuracy | correlation: r = 0.647 | Improved |
| evaluate mm-TEM vs Titans and Transformer under counterclockwise shift | mm-TEM · circular-grid test | accuracy: over 90% vs drops of up to 30% | Improved |
Closest papers by meaning
- Binding in hippocampal-entorhinal circuits enables compositionality in cognitive maps (NeurIPS 2024 · similarity 0.87)
- A Multi-Region Brain Model to Elucidate the Role of Hippocampus in Spatially Embedded Decision-Making (ICML 2025 · similarity 0.87)
- Leveraging Attractor Dynamics in Spatial Navigation for Better Language Parsing (ICML 2024 · similarity 0.86)
- Place Cells as Multi-Scale Position Embeddings: Random Walk Transition Kernels for Path Planning (NeurIPS 2025 · similarity 0.85)
- Spatially-Aware Transformers for Embodied Agents (ICLR 2024 · similarity 0.85)
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.85)
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.
