Research map / Vision-language and multimodal / Visual question answering
How Modular should Neural Module Networks Be for Systematic Generalization?
Read the paperCheck 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 group-level image encoder modules instead of a single shared encoder (ablation) | Vector-NMN · CLEVR-CoGenT | systematic generalization accuracy (%): 73.2 -> 77.3 | Helped |
| use group-level image encoder modules instead of a single shared encoder (ablation) | Vector-NMN · CLEVR-CoGenT in-distribution | accuracy (%): 98.0 -> 94.4 | Hurt |
| group image encoder (ablation) | Vector-NMN · CLEVR-CoGenT query_shape | systematic generalization accuracy (%): 40 -> 52 | Helped |
| group image encoder (ablation) | Vector-NMN · CLEVR-CoGenT query_color | systematic generalization accuracy (%): 60 -> 67 | Helped |
| group image encoder (ablation) | Vector-NMN · CLEVR-CoGenT equal_shape | systematic generalization accuracy (%): 89 -> 79 | Hurt |
| group image encoder (ablation) | Vector-NMN · CLEVR-CoGenT exist | systematic generalization accuracy (%): 84.4 -> 84.4 | No clear effect |
| use sub-task image-encoder library instead of all-all-all on SQOOP with two objects per image | NMN · SQOOP 2 objects | systematic generalization accuracy (%): 84 -> 88.5 | Helped |
| use sub-task intermediate modules instead of all-all-all on SQOOP with five objects per image | NMN · SQOOP 5 objects | systematic generalization accuracy (%): 99.8 -> 99.96 | No clear effect |
| use Vector-NMN instead of Tensor-NMN | NMN with all encoder · CLEVR-CoGenT | systematic generalization accuracy (%): 72.7 -> 73.2 | No clear effect |
| intermediate degree of modularity (group) at image encoder instead of none or full sub-task modularity (ablation) | NMN libraries · VQA-MNIST (4 tasks) | systematic generalization accuracy (%) | Helped |
| full sub-task modularity at all stages instead of group (ablation) | NMN libraries · VQA-MNIST | systematic generalization accuracy (%) | Hurt |
| modularity at the classifier stage only (ablation) | NMN · VQA-MNIST | systematic generalization accuracy (%) | No clear effect |
Closest papers by meaning
- Breaking Neural Network Scaling Laws with Modularity (ICLR 2025 · similarity 0.90)
- Neural Module Networks (CVPR 2016 · similarity 0.89)
- Detection-Based Intermediate Supervision for Visual Question Answering (AAAI 2024 · similarity 0.86)
- GLEN: Generalized Focal Loss Ensemble of Low-Rank Networks for Calibrated Visual Question Answering (AAAI 2025 · similarity 0.85)
- Exploring the Effectiveness of Object-Centric Representations in Visual Question Answering: Comparative Insights with Foundation Models (ICLR 2025 · similarity 0.85)
- Neural networks are more modular than single neurons suggest (NeurIPS 2026 · similarity 0.85)
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.