Research map / Vision-language and multimodal / Visual question answering
Learning to Specialize with Knowledge Distillation for Visual Question Answering
Jonghwan Mun, Kimin Lee, Jinwoo Shin, Bohyung Han
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add knowledge distillation loss for non-assigned models in MCL (MCL-KD vs MCL) (ablation) | MCL ensemble · CLEVR val, SAN, k=1 | top-1 accuracy: 42.19 -> 85.98 | Improved |
| replace uniform-distribution regularization (CMCL) with knowledge distillation (MCL-KD) (ablation) | CMCL · CLEVR val, SAN, k=3 | top-1 accuracy: 86.18 -> 88.16 | Improved |
| MCL-KD instead of independent ensemble | independent ensemble · CLEVR val, SAN, k=1 | top-1 accuracy: 85.23 -> 85.98 | Improved |
| MCL-KD instead of independent ensemble | independent ensemble · CLEVR val, MLP, k=1 | top-1 accuracy: 60.10 -> 60.22 | Improved |
| MCL-KD instead of independent ensemble | independent ensemble · VQA v2.0 val, k=3 | top-1 accuracy: 65.27 -> 65.67 | Improved |
| CMCL instead of independent ensemble | independent ensemble · VQA v2.0 val, k=3 | top-1 accuracy: 65.27 -> 64.99 | No clear effect |
| increase number of specialized models per example k from 1 to 3 in MCL-KD (ablation) | MCL-KD · CLEVR val, SAN | top-1 accuracy: 85.98 -> 88.16 | Improved |
| MCL-KD instead of independent ensemble | independent ensemble · CIFAR-100, ResNet-20, k=1 | top-1 accuracy: 64.60 -> 65.27 | Improved |
| MCL-KD instead of independent ensemble | independent ensemble · CIFAR-100, VGGNet-17, k=1 | top-1 accuracy: 68.43 -> 68.75 | Improved |
| MCL and CMCL instead of independent ensemble | independent ensemble · CIFAR-100, ResNet-20, k=1 | top-1 accuracy: 64.60 -> 56.40 (MCL), 48.67 (CMCL) | Hurt |
| add feature sharing to MCL-KD (ablation) | MCL-KD · CIFAR-100, ResNet-20, k=1 | top-1 accuracy: 65.27 -> 66.72 | Improved |
| add feature sharing to CMCL (ablation) | CMCL · CIFAR-100, ResNet-20, k=1 | top-1 accuracy: 48.67 -> 56.30 | Improved |
Closest papers by meaning
- Plenty is Plague: Fine-Grained Learning for Visual Question Answering (TPAMI 2019 · similarity 0.91)
- Multi-Question Learning for Visual Question Answering (AAAI 2020 · similarity 0.91)
- Visual Question Answering as a Meta Learning Task (ECCV 2018 · similarity 0.91)
- Separating Skills and Concepts for Novel Visual Question Answering (CVPR 2021 · similarity 0.90)
- Revisiting Visual Question Answering Baselines (ECCV 2016 · similarity 0.90)
- Multimodal Commonsense Knowledge Distillation for Visual Question Answering (Student Abstract) (AAAI 2025 · similarity 0.89)
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.
