Research map / Architectures and efficiency / Attention and transformers
Meshed-Memory Transformer for Image Captioning
Check 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 3 layers instead of 6 in the Transformer (ablation) | Transformer · COCO Karpathy test split | CIDEr: 121.8 -> 123.6 | Helped |
| Add Attention on Attention (ablation) | 3-layer Transformer · COCO Karpathy test split | CIDEr: 123.6 -> 129.1 | Helped |
| Replace last-layer-only cross-attention with 1-to-1 encoder-decoder connectivity (ablation) | Transformer · COCO Karpathy test split | CIDEr: 123.6 -> 129.2 | Helped |
| Use meshed connectivity across all encoder layers with sigmoid gating (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 129.2 -> 131.2 | Helped |
| Add memory vectors to the 1-to-1 model (ablation) | M² Transformer 1-to-1 · COCO Karpathy test split | CIDEr: 128.4 -> 129.2 | Helped |
| Add memory vectors (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 129.4 -> 131.2 | Helped |
| Replace sigmoid gating with softmax gating (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 131.2 -> 130.3 | Hurt |
| Increase memory vectors from 20 to 40 (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 129.9 -> 131.2 | Helped |
| Increase memory vectors from 40 to 60 (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 131.2 -> 129.6 | Hurt |
| Increase memory vectors from 40 to 80 (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 131.2 -> 128.9 | Hurt |
| Use 2 instead of 3 encoder and decoder layers (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 131.2 -> 128.5 | Hurt |
| Use 4 instead of 3 encoder and decoder layers (ablation) | M² Transformer · COCO Karpathy test split | CIDEr: 131.2 -> 129.6 | Hurt |
Closest papers by meaning
- Frozen Transformers in Language Models Are Effective Visual Encoder Layers (ICLR 2024 · similarity 0.87)
- Towards General Continuous Memory for Vision-Language Models (NeurIPS 2025 · similarity 0.87)
- Memory-Augmenting Decoder-Only Language Models through Encoders (Student Abstract) (AAAI 2024 · similarity 0.86)
- MemoryFormer : Minimize Transformer Computation by Removing Fully-Connected Layers (NeurIPS 2024 · similarity 0.86)
- Exploring the Role of Large Language Models in Prompt Encoding for Diffusion Models (NeurIPS 2024 · similarity 0.86)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.86)
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.