Research map / Vision-language and multimodal / Human and embodied modeling
Interpretable Visual Reasoning via Induced Symbolic Space
Zhonghao Wang, Kai Wang, Mo Yu, Jinjun Xiong, Wen-mei Hwu, Mark Hasegawa-Johnson, Humphrey Shi
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 |
|---|---|---|---|
| increase MAC reasoning steps from 4 to 8 (ablation) | OCCAM · CLEVR | accuracy: 94.3 -> 98.6 | Improved |
| increase MAC reasoning steps from 8 to 12 (ablation) | OCCAM · CLEVR | accuracy: 98.6 -> 99.4 | Improved |
| increase MAC reasoning steps from 12 to 16 (ablation) | OCCAM · CLEVR | accuracy: 99.4 -> 99.1 | No clear effect |
| increase MAC reasoning steps from 4 to 16 (ablation) | OCCAM · GQA test-dev | accuracy: 55.1 -> 56.2 | Improved |
| use learnable ResNet34 object features instead of pretrained ResNet101 (ablation) | OCCAM · CLEVR val | accuracy: 97.9% -> 99.0% | Improved |
| add global features to learnable ResNet34 object features (ablation) | OCCAM · CLEVR val | accuracy: 99.0% -> 99.4% | Improved |
| constrain attention to object-level features instead of image-level attention | MAC · CLEVR and GQA validation sets | accuracy | Improved |
| replace visual features with induced symbolic concepts | OCCAM · CLEVR and GQA | accuracy | No clear effect |
| use induced concept correlations instead of word2vec embeddings for synonym classification | word concept grouping · Human study synonyms/uncorrelated words | overall accuracy: 68.37% -> 78.06% | Improved |
| use induced concept proximity instead of GloVe embeddings for word ranking | word concept grouping · Word proximity ranking vs human | Kendall tau ranking distance: 0.3418 -> 0.2585 | Improved |
Closest papers by meaning
- The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision (ICLR 2019 · similarity 0.89)
- Transparency by Design: Closing the Gap Between Performance and Interpretability in Visual Reasoning (CVPR 2018 · similarity 0.89)
- Neuro-Symbolic Visual Reasoning: Disentangling "Visual" from "Reasoning" (ICML 2020 · similarity 0.89)
- Exploring the Effectiveness of Object-Centric Representations in Visual Question Answering: Comparative Insights with Foundation Models (ICLR 2025 · similarity 0.88)
- Visual Chain-of-Thought Prompting for Knowledge-Based Visual Reasoning (AAAI 2024 · similarity 0.88)
- Visual Graph Arena: Evaluating Visual Conceptualization of Vision and Multimodal Large Language Models (ICML 2025 · similarity 0.87)
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.
