Research map / Label-efficient and robust learning / Self-supervised and contrastive

VCR: Learning Valid Contextual Representation for Incomplete Wearable Signals

NeurIPS 2026 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Invariance-based self-supervised learning · 0 citations (OpenAlex)

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 VCR, a self-supervised wearable-signal pretraining framework that uses a geometric orthogonal tokenizer to split each modality into shared and specific tokens and a missing-aware MoE backbone that reconstructs only inferable shared components of missing modalities.

Problem

Wearable health models need label-efficient pretraining that stays robust when modalities are missing at test time, but reconstruction-based methods hallucinate non-inferable modality-specific details when a modality is absent.

Approach

Strict geometric orthogonal projection (QR-parameterized basis plus residual adapter and whitening) separates each modality into shared and specific components, and the objective reconstructs only the shared component of a single missing modality while skipping specific components, with a sparse MoE backbone that routes tokens from observed modalities only.</ANSWER_PLACEHOLDER>

Key results

• On full-modality WESAD, VCR (FT) reaches Macro-F1 0.707 vs 0.610 for the best supervised baseline (ResNet).
• On WEEE regression, VCR (FT) MAE 2.633 vs 3.039 for FlexMoE under full modality.
• When ACC is missing on DaLiA, VCR (FT) maintains F1 0.416 and VCR (LP) 0.428, versus lower scores for baselines.

Surprising findings

• Reconstructing the shared component of a missing modality is kept in the objective even though it is partly predictable from observed modalities, and this improves robustness.
• Cross-modal analysis shows that per-sample representation shift from reconstruction correlates with class probability drop (r=0.561) for LSM-2.

Limitations

• Evaluation limited to a single wearable device (Empatica E4), so cross-device generalization is untested.
• Pretraining and downstream results for baselines are reported with ~50M capacity-matched configurations; generalization beyond these four tasks is unverified.

Experimental setup

Datasets: BIG IDEAs UE4W CAN-STRESS WESAD DaLiA AAUWSS WEEE

Benchmarks: WESAD DaLiA AAUWSS WEEE

Metrics: Macro-F1 Balanced Accuracy Cohen's kappa MAE R^2 HSIC

Baselines: ResNet ViT FlexMoE SimCLR MSN CMC MF-CLR FOCAL LSM LSM-2

Builds on: FOCAL LSM-2 FlexMoE masked autoencoding

Closest papers by meaning

• Mitigating Pervasive Modality Absence Through Multimodal Generalization and Refinement (AAAI 2025 · similarity 0.85)
• SensorLM: Learning the Language of Wearable Sensors (NeurIPS 2025 · similarity 0.85)
• X-Fi: A Modality-Invariant Foundation Model for Multimodal Human Sensing (ICLR 2025 · similarity 0.84)
• Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.84)
• Flex-MoE: Modeling Arbitrary Modality Combination via the Flexible Mixture-of-Experts (NeurIPS 2024 · similarity 0.84)
• Do More Modalities Always Help? A Geometric Perspective on Missing-Modality Robustness (NeurIPS 2026 · similarity 0.84)