Research map / Vision-language and multimodal / Human and embodied modeling
data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use L1 instead of L2 pre-training loss (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 17.2 | No clear effect |
| Use Smooth L1 loss with β=0.08 instead of L2 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 17.2 | No clear effect |
| Use Smooth L1 loss with β=0.25 instead of L2 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 16.8 | Helped |
| Use Smooth L1 loss with β=0.5 instead of L2 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 16.8 | Helped |
| Use Smooth L1 loss with β=1 instead of L2 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 17.3 | Hurt |
| Raise speech mask probability to 0.8 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 17.2 | No clear effect |
| Lower speech mask probability to 0.5 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 17.3 | No clear effect |
| Shorten speech mask length to 5 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 22.1 | Hurt |
| Lengthen speech mask to 15 (ablation) | data2vec speech · Librispeech dev-other | WER: 17.1 -> 18.9 | Hurt |
| Average multiple teacher layers for targets instead of using only the top layer (ablation) | data2vec · ImageNet; Libri-light 10-hour dev-other; GLUE validation | top-1 accuracy; WER; average GLUE score | Helped |
| Give teacher targets access to a larger input context, up to the entire sample (ablation) | data2vec · ImageNet; Libri-light 10-hour dev-other; GLUE validation | top-1 accuracy; WER; average GLUE score | Helped |
| Use FFN output features rather than self-attention output features as speech targets (ablation) | data2vec speech · Librispeech | downstream performance | Helped |
Closest papers by meaning
- Efficient Self-supervised Learning with Contextualized Target Representations for Vision, Speech and Language (ICML 2023 · similarity 0.94)
- wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (NeurIPS 2020 · similarity 0.92)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.87)
- Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Word2Vec-like Models (NeurIPS 2025 · similarity 0.86)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (NeurIPS 2024 · similarity 0.85)
- ES3: Evolving Self-Supervised Learning of Robust Audio-Visual Speech Representations (CVPR 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.