Research map / Label-efficient and robust learning / Self-supervised and contrastive
MICE: Multi-animal Interaction Context Encoder — A Hierarchical Foundation Model for Mouse Behavior
Yen-Shuo Su, Weihan Li, Anqi Wu
Read the paperCheck your idea against 137,217 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 MICE, a hierarchical self-supervised foundation model for mouse behavior from pose that encodes multi-animal social interactions.
Problem
Self-supervised pose models for mouse behavior encode each animal independently and thus fail to capture inherently multi-animal social behaviors such as aggression, courtship, mounting, and allogrooming.
Approach
A hierarchical encoder over multi-animal pose that models interaction context across animals, rather than encoding each animal in isolation, as a foundation model for mouse behavior.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add social encoder output to individual embedding (ablation) | MICE · AdaptableSnail (combined) | macro-F1 (%): 22.66 -> 26.76 | Improved |
| Add social encoder output to individual embedding (ablation) | MICE · MABe22 (combined) | macro-F1 (%): 67.85 -> 69.26 | Improved |
| Replace correct partners with mis-paired partners when training the social stage (ablation) | MICE social stage · AdaptableSnail | macro-F1 (%): 26.76 -> 22.79 | Hurt |
| Remove partners from social stage (ablation) | MICE social stage · CalMS21 | macro-F1 (%): 84.56 -> 81.00 | Hurt |
| Joint pretraining on six datasets instead of single-dataset | MICE · AdaptableSnail | macro-F1 (%): 21.93 -> 26.76 | Improved |
| Joint pretraining on six datasets instead of single-dataset | MICE · MouseLockbox | macro-F1 (%): 15.93 -> 13.68 | Hurt |
| Leave the target dataset out of pretraining (ablation) | MICE · MABe22 | macro-F1 (%): 69.26 -> 64.28 | Hurt |
| Handle each mouse separately (MICE ind_cat) instead of stacking N animals (Progressive) (ablation) | MICE individual encoder · MABe22 (combined) | macro-F1 (%): 61.87 -> 67.85 | Improved |
| Widen attention windows across stages instead of multi-resolution supervision only (Multiscale) (ablation) | MICE individual encoder · MABe22 (combined) | macro-F1 (%): 63.68 -> 67.85 | Improved |
| Use MICE instead of hBehaveMAE | unsupervised segmentation · 4 datasets avg | ARI (%): 31.32 -> 43.33 | Improved |
| Use MICE instead of hBehaveMAE | unsupervised segmentation · 4 datasets avg | Boundary-F1: 31.37 -> 31.57 | No clear effect |
Closest papers by meaning
- The Multi-Agent Behavior Dataset: Mouse Dyadic Social Interactions (NeurIPS 2021 · similarity 0.92)
- Of Mice and Mates: Automated Classification and Modelling of Mouse Behaviour in Groups Using a Single Model Across Cages (IJCV 2024 · similarity 0.92)
- Mouse Lockbox Dataset: Behavior Recognition for Mice Solving Lockboxes (IJCV 2026 · similarity 0.91)
- Pose Splatter: A 3D Gaussian Splatting Model for Quantifying Animal Pose and Appearance (NeurIPS 2025 · similarity 0.86)
- Inpainting the Neural Picture: Inferring Unrecorded Brain Area Dynamics from Multi-Animal Datasets (NeurIPS 2025 · similarity 0.85)
- Animal-Bench: Benchmarking Multimodal Video Models for Animal-centric Video Understanding (NeurIPS 2024 · similarity 0.84)
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.
