Research map / Label-efficient and robust learning / Recognition and long-tail
ReMAP: Neural Model Reprogramming with Network Inversion and Retrieval-Augmented Mapping for Adaptive Motion Forecasting
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 ReMAP, a model reprogramming method that maps impaired-limb gait inputs to able-bodied correction templates (via retrieval or network inversion) so a frozen able-bodied gait forecasting model can predict ankle motion for below-knee impaired individuals.
Problem
Forecasting reference joint motion for limb-impaired patients is hard because joint locomotion data from impaired individuals is scarce, and models trained on able-bodied gait transfer poorly.
Approach
Keep an able-bodied multi-task gait model frozen and train a small MLP refurbish module that maps impaired inputs to correction templates found by retrieval-augmented nearest-neighbor search or network inversion, optionally combined with a target-based loss (hybrid), so the frozen model outputs the desired impaired-limb motion.
Key results
- With a train sample ratio of 0.1, hybrid (inversion) reached R2 0.87±0.09 (task-specific heads) and hybrid (neighbor) 0.86±0.07, versus 0.74±0.22 for fine-tuning and 0.71±0.14 for direct mapping.
- Cross-mapping zero-shot transfer gave negative R2 (-0.35 and -0.32), showing able-bodied models transfer poorly to impaired gait.
- Hybrid strategies gave roughly 11-13% improvement over fine-tuning in the low-data regime, per the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine correction-based and target-based mapping with hybrid (neighbor) | Task-shared ReMAP · Impaired motion prediction, train sample ratio 0.1 | R2: 0.77 ± 0.2 -> 0.84 ± 0.07 | Helped |
| combine correction-based and target-based mapping with hybrid (neighbor) | Task-specific ReMAP · Impaired motion prediction, train sample ratio 0.1 | R2: 0.71 ± 0.14 -> 0.86 ± 0.07 | Helped |
| combine correction-based and target-based mapping with hybrid (inversion) | Task-specific ReMAP · Impaired motion prediction, train sample ratio 0.1 | R2: 0.74 ± 0.22 -> 0.87 ± 0.09 | Helped |
| use target-based mapping without the correction-based loss (ablation) | ReMAP · Impaired motion prediction, train sample ratio 0.1, task-shared | R2: 0.84 ± 0.07 -> 0.39 ± 0.25 | Hurt |
| use target-based mapping without the correction-based loss (ablation) | ReMAP · Impaired motion prediction, train sample ratio 0.1, task-specific | R2: 0.86 ± 0.07 -> 0.61 ± 0.22 | Hurt |
| use correction-based mapping without target-based loss (ablation) | ReMAP · Impaired motion prediction, train sample ratio 0.1, task-shared | R2: 0.84 ± 0.07 -> 0.33 ± 0.11 | Hurt |
| use correction-based mapping without target-based loss (ablation) | ReMAP · Impaired motion prediction, train sample ratio 0.1, task-specific | R2: 0.86 ± 0.07 -> 0.19 ± 0.22 | Hurt |
| use network inversion without target-based loss (ablation) | ReMAP · Impaired motion prediction, train sample ratio 0.1, task-shared | R2: 0.83 ± 0.09 -> 0.23 ± 0.07 | Hurt |
| use network inversion without target-based loss (ablation) | ReMAP · Impaired motion prediction, train sample ratio 0.1, task-specific | R2: 0.87 ± 0.09 -> 0.39 ± 0.25 | Hurt |
| increase the target-based loss weight β from 1 to 20 (ablation) | Hybrid ReMAP · Impaired motion prediction | R2: 0.63±0.04 -> 0.94±0.01 | Helped |
| increase the target-based loss weight β from 20 to 50 (ablation) | Hybrid ReMAP · Impaired motion prediction | R2: 0.94±0.01 -> 0.85±0.07 | Hurt |
| increase correction-based training data from train size 0.1 to 0.4 (ablation) | Correction-based mapping · Impaired motion prediction | Output R2: 0.18±0.14 -> 0.58±0.11 | Helped |
Surprising findings
- Correction-based mapping alone performs poorly with little data (R2 around 0.18 at 10% training), but adding a target-based loss makes the hybrid strategy strongest.
Limitations
- Ground-truth ankle trajectories for impaired subjects are approximated from anthropometry- and speed-matched able-bodied subjects, so suitability for powered prosthesis control needs validation.
- Data is restricted by a consortium agreement, so the study cannot be readily reproduced.
- Evaluated only on below-knee impaired individuals and a small number of subjects.
Experimental setup
Datasets: able-bodied locomotion wearable-sensor dataset (public, CC BY 4.0) below-knee lower-limb impairment motion capture dataset (Vicon, OpenSim-derived joint angles)
Benchmarks: below-knee impaired individuals gait dataset (in-house)
Metrics: coefficient of determination (R2)
Baselines: cross-mapping (zero-shot transfer) direct mapping fine-tuning (transfer learning)
Builds on: model reprogramming network inversion (Linden and Kindermann) retrieval-augmented generation
Closest papers by meaning
- Care-PD: A Multi-Site Anonymized Clinical Dataset for Parkinson’s Disease Gait Assessment (NeurIPS 2025 · similarity 0.83)
- VidSole: A Multimodal Dataset for Joint Kinetics Quantification and Disease Detection with Deep Learning (AAAI 2025 · similarity 0.83)
- RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data (ICLR 2025 · similarity 0.82)
- Exploring the trade-off between deep-learning and explainable models for brain-machine interfaces (NeurIPS 2024 · similarity 0.82)
- Motif Guided Graph Transformers with Combinatorial Skeleton Prototype Learning for Skeleton-Based Person Re-Identification (AAAI 2025 · similarity 0.81)
- SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · similarity 0.81)
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.