Reproduce “GAUGE: Granularity-Adaptive Counterfactual Gating of Evidence for Incomplete Multimodal Classification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunping Shi, En Yu, Kairui Guo, Jie Lu
GAUGE, a framework from the Australian Artificial Intelligence Institute (UTS), introduces granularity-adaptive counterfactual gating to enhance incomplete multimodal classification by selectively modulating individual evidence units within both observed and reconstructed modalities. It consistently outperformed existing methods across six benchmarks, achieving up to 5.90 percentage points accuracy improvement over modality-level control in severe missingness scenarios.
Key insights
- Effective handling of incomplete multimodal data requires control at a fine-grained, unit-level rather than coarse modality-level, especially when dealing with potentially unreliable reconstructed information.
- Prediction-aware counterfactual scoring, efficiently approximated by a first-order Taylor expansion, provides a principled and scalable method to assess the contribution of individual evidence units to model predictions.
- Modulating attention logits via additive biases based on these scores offers a non-invasive yet powerful way to integrate granularity-adaptive evidence gating into existing Transformer-based architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.