AWhat changed
Regressed contextual latent representations from an EMA teacher's full input using masked student views, applying one self-distillation objective across separate speech, image and text settings. data2vec2 skipped masked inputs, used a convolutional decoder and amortized teacher computation across several masked student views.
Scientific scope
Regressed contextual latent representations from an EMA teacher's full input using masked student views, applying one self-distillation objective across separate speech, image and text settings. data2vec2 skipped masked inputs, used a convolutional decoder and amortized teacher computation across several masked student views. Inherits EMA teachers, BYOL/DINO-style self-distillation and masking. Same learning method across modalities is not a single jointly trained multimodal model: feature encoders and masking remain modality specific. Not invention of all latent prediction or evidence for a biological learning theory; benchmark gains and anti-collapse behavior have configuration limits. Version2 inherits visible-only encoding and multi-mask reuse; reported16.4fold image pretraining reduction is protocol-specific, not every modality or end-to-end workflow.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits EMA teachers, BYOL/DINO-style self-distillation and masking. Same learning method across modalities is not a single jointly trained multimodal model: feature encoders and masking remain modality specific. Not invention of all latent prediction or evidence for a biological learning theory; benchmark gains and anti-collapse behavior have configuration limits. Version2 inherits visible-only encoding and multi-mask reuse; reported16.4fold image pretraining reduction is protocol-specific, not every modality or end-to-end workflow.
Contributors
- Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
- data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language
- Efficient Self-supervised Learning with Contextualized Target Representations for Vision, Speech and Language
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction