Research map / Vision-language and multimodal / Visual question answering
Removing Bias in Multi-modal Classifiers: Regularization by Maximizing Functional Entropies
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 |
|---|---|---|---|
| Add tensorized functional Fisher information regularization | LMH · VQA-CPv2 test | Overall accuracy: 52.05 -> 54.55 | Helped |
| Add functional Fisher information regularization | LMH · VQA-CPv2 test, Number answers | Accuracy: 31.12 -> 49.16 | Helped |
| Add functional Fisher information regularization | LMH · VQA-CPv2 test, Other answers | Accuracy: 46.97 -> 45.82 | Hurt |
| Add tensorized functional Fisher information regularization | SocialIQ classifier · SocialIQ validation | Accuracy: 63.91±0.26 -> 67.93±0.18 | Helped |
| Add tensorized functional Fisher information regularization | Colored MNIST classifier · Colored MNIST grayscale validation | Accuracy after convergence: 41.11±2.13 -> 96.17±0.63 | Helped |
| Add tensorized functional Fisher information regularization | Image classifier · Dogs & Cats TB1 | Accuracy after convergence: 79.22±0.45 -> 94.71±0.37 | Helped |
| Add tensorized functional Fisher information regularization | Image classifier · Dogs & Cats TB2 | Accuracy after convergence: 65.51±1.54 -> 88.11±0.17 | Helped |
| Use non-tensorized rather than tensorized functional Fisher regularization (ablation) | Proposed regularized classifier · SocialIQ | Accuracy after convergence: 67.93±0.18 -> 63.87±0.34 | Hurt |
| Use functional entropy regularization rather than the Fisher-information-bound variant (ablation) | Proposed regularized classifier · SocialIQ | Accuracy after convergence: 67.93±0.18 -> 64.36±0.31 | Hurt |
| Use variance regularization instead of functional entropy regularization (ablation) | Proposed regularized classifier · SocialIQ | Accuracy after convergence: 67.93±0.18 -> 67.41±0.21 | Hurt |
| Use non-tensorized rather than tensorized functional Fisher regularization (ablation) | Proposed regularized classifier · Colored MNIST grayscale validation | Accuracy after convergence: 96.17±0.63 -> 93.68±0.75 | Hurt |
| Use functional entropy regularization rather than the Fisher-information-bound variant (ablation) | Proposed regularized classifier · Colored MNIST grayscale validation | Accuracy after convergence: 96.17±0.63 -> 94.87±1.03 | Hurt |
Closest papers by meaning
- Functional Regularization for Representation Learning: A Unified Theoretical Perspective (NeurIPS 2020 · similarity 0.86)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.85)
- Cross-Modal Feature Distribution Calibration for Few-Shot Visual Question Answering (AAAI 2024 · similarity 0.85)
- MultiMax: Sparse and Multi-Modal Attention Learning (ICML 2024 · similarity 0.85)
- Bidirectional Contrastive Split Learning for Visual Question Answering (AAAI 2024 · similarity 0.85)
- Directional Gradient Projection for Robust Fine-Tuning of Foundation Models (ICLR 2025 · similarity 0.83)
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.