Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning to Draw: Emergent Communication through Sketching
Daniela Mihai, Jonathon Hare
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 a perceptual loss matching sketch and photo features | Drawing communication agents · STL-10 original game; human receiver with 9 distractors | human communication rate: 8.3\%\ (\pm5.4) -> 38.3\%\ (\pm2.5) | Improved |
| Add perceptual loss | Drawing communication agents · STL-10 original game | communication success rate: 71.8\%\ (\pm 6.1) -> 69.57\%\ (\pm 2.6) | No clear effect |
| Use an object-oriented game with disjoint classes | Drawing communication agents · STL-10 | human interpretability | Improved |
| Increase sender encoding and receiver MLP width from 64 to 1024 (ablation) | Drawing communication agents · Caltech-101 OO-game different | communication rate: 50.46\%\ (\pm 1.5) -> 64.99\%\ (\pm 1.5) | Improved |
| Replace ImageNet-pretrained weights with Stylized-ImageNet shape-biased weights (ablation) | Drawing communication agents · Caltech-101 original game | communication success rate: 78.46\%\ (\pm 2.0) -> 77.09\%\ (\pm 1.9) | No clear effect |
| Reduce the sketch from 20 strokes to 5 (ablation) | Drawing communication agents · STL-10 original game | communication success rate: 69.57\% \ (\pm 2.6) -> 73.41\%\ (\pm 1.6) | Improved |
| Increase stroke count from 20 to 50 (ablation) | Drawing communication agents · STL-10 original game | communication success rate: 69.57\% \ (\pm 2.6) -> 69.21\% \ (\pm 1.4) | Hurt |
| Use only VGG16 layers 1–4 rather than all five feature layers in perceptual loss (ablation) | Drawing communication agents · STL-10 original game | communication success rate: 69.57\%\ (\pm 2.6) -> 19.09\%\ (\pm 10.1) | Hurt |
| Use perceptual features only up to VGG16 layer 3 (ablation) | Drawing communication agents · STL-10 original game | communication success rate: 69.57\%\ (\pm 2.6) -> 1.0\% \ (\pm 0) | Hurt |
| Use only the first or second VGG16 feature block for perceptual loss (ablation) | Drawing communication agents · STL-10 original game | human interpretability | Hurt |
| Increase the number of perceptual feature maps beyond the original five (ablation) | Drawing communication agents · STL-10 original game | communication success rate: 68.4\%\ (\pm 2.0); 69.5\%\ (\pm 1.8); 69.0\% \ (\pm 1.2) | No clear effect |
| Replace photo-matching perceptual loss with an arbitrary dog-image objective (ablation) | Drawing communication agents · CelebA original game | human interpretability | Hurt |
Closest papers by meaning
- Emergent Graphical Conventions in a Visual Communication Game (NeurIPS 2022 · similarity 0.94)
- SketchAgent: Language-Driven Sequential Sketch Generation (CVPR 2025 · similarity 0.88)
- Learning Multi-Object Positional Relationships via Emergent Communication (AAAI 2024 · similarity 0.88)
- Sketchtopia: A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback (CVPR 2025 · similarity 0.87)
- Speaking Your Language: Spatial Relationships in Interpretable Emergent Communication (NeurIPS 2024 · similarity 0.86)
- A compressive-expressive communication framework for compositional representations (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.
