Research map / Architectures and efficiency / Mixture of experts
MoE-ViE: Mixture of Experts Vision Encoder for Efficient Image and Video Understanding
Bonan Zhang, Shiyu Dong, Quan Tran, Katharina Gschwind, Shuqi Yang, Sijia Chen, Adel Ahmadyan, Seungwhan Moon, Lu Zhang, Ahmed Kirmani, Babak Damavandi, Anuj Kumar
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 |
|---|---|---|---|
| replace dense MLP with fine-grained MoE (ViT-B/32) (ablation) | CLIP vision encoder · Zero-shot avg classification (B/32, 12.8B samples) | avg accuracy: 55.8 -> 63.2 | Improved |
| use fine-grained MoE instead of conventional MoE (ViT-B/32) (ablation) | MoE-ViE · Zero-shot avg classification (B/32, 12.8B samples) | avg accuracy: 59.9 -> 63.2 | Improved |
| use fine-grained MoE instead of conventional MoE (ViT-L/16) (ablation) | MoE-ViE · Zero-shot avg classification (L/16, 12.8B samples) | avg accuracy: 78.5 -> 79.6 | Improved |
| z-score magnitude-aware bias update instead of sign update in loss-free balancing (ablation) | MoE-ViE-B/32 · Load balancing ablation (MoE-ViE-B/32) | avg classification: 62.6 -> 63.2 | Improved |
| loss-free balancing instead of importance/load auxiliary loss (ablation) | MoE-ViE-B/32 · Load balancing ablation (MoE-ViE-B/32) | avg classification: 59.9 -> 62.6 | Improved |
| add one shared expert (ablation) | MoE-ViE-B/32 · Shared experts ablation (MoE-ViE-B/32) | avg classification: 62.1 -> 63.2 | Improved |
| increase shared experts from 1 to 4 (ablation) | MoE-ViE-B/32 · Shared experts ablation (MoE-ViE-B/32) | avg classification: 63.2 -> 62.5 | Hurt |
| vary shared expert weight lambda from 0.01 to 10 (ablation) | MoE-ViE-B/32 · Weight of shared experts | avg classification: 62.9 - 63.2 | No clear effect |
| increase top-k from 8 to 16 (ablation) | MoE-ViE-B/32 · Top-k ablation (MoE-ViE-B/32) | avg classification: 63.2 -> 63.1 | No clear effect |
| reduce top-k from 8 to 4 (ablation) | MoE-ViE-B/32 · Top-k ablation (MoE-ViE-B/32) | avg classification: 63.2 -> 62.4 | Hurt |
| make experts finer (width 1/4 vs full width) (ablation) | MoE-ViE-B/32 · Expert width ablation (MoE-ViE-B/32) | avg classification: 59.9 -> 63.2 | Improved |
| make experts finer still (1/8 width vs 1/4) (ablation) | MoE-ViE-B/32 · Expert width ablation (MoE-ViE-B/32) | avg classification: 63.2 -> 62.8 | No clear effect |
Closest papers by meaning
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.89)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.89)
- Mixture of Nested Experts: Adaptive Processing of Visual Tokens (NeurIPS 2024 · similarity 0.89)
- MoVA: Adapting Mixture of Vision Experts to Multimodal Context (NeurIPS 2024 · similarity 0.88)
- MoE Jetpack: From Dense Checkpoints to Adaptive Mixture of Experts for Vision Tasks (NeurIPS 2024 · similarity 0.88)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.88)
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.
