AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

VidLanKD: Improving Language Understanding via Video-Distilled Knowledge Transfer

Zineng Tang, Jaemin Cho, Hao Tan, Mohit Bansal

NeurIPS 2021 · Vision-language and multimodal › Human and embodied modeling · approach: Vision-language pretraining · 3 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (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 changedWhereMeasuredEffect
Distill with video-grounded NST+CRD instead of text-only pretrainingBERT-base · 7 downstream NLU tasks (Avg.)average score: 78.9 -> 82.6Improved
Use NST+CRD instead of image-based vokenizationBERT-base · 7 downstream NLU tasks (Avg.)average score: 81.4 -> 82.6Improved
Use a CLIP frame encoder instead of ResNet (ablation)VidLanKD · 7 downstream NLU tasks (Avg.)average score: 82.4 -> 82.6Improved
Use an MLM-only teacher instead of the text-only baseline (ablation)BERT-small with NST distillation · SST-2accuracy: 88.8 -> 88.1Hurt
Use a contrastive-only teacher instead of the text-only baseline (ablation)BERT-small with NST distillation · QNLIaccuracy: 84.9 -> 85.2Improved
Train the teacher with both MLM and contrastive objectives (ablation)BERT-small with NST distillation · SST-2accuracy: 88.8 -> 91.1Improved
Use video-then-text two-stage pretraining instead of VidLanKD (ablation)BERT-small · SST-2accuracy: 91.1 -> 90.3Hurt
Use soft-label distillation instead of NST+CRD (ablation)BERT-small · SST-2accuracy: 91.5 -> 87.2Hurt
Combine NST and CRD rather than using NST alone (ablation)BERT-small · SST-2accuracy: 91.1 -> 91.5Improved
Use video-level NST rather than video vokenization (ablation)BERT-small · SST-2accuracy: 89.7 -> 91.1Improved
Use video-level rather than image-level NST supervision (ablation)BERT-small · SST-2accuracy: 88.9 -> 91.1Improved
Distill video-grounded knowledge with NSTBERT-small · GLUE diagnostics Knowledgescore: +4.6Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy