Why Can't I Open My Drawer? Mitigating Object-Driven Shortcuts in Zero-Shot Compositional Action Recognition

1. 1Identify sparse compositional supervision as a potential cause of object-driven shortcuts
2. 2Recognize verb-object learning asymmetry as a factor contributing to shortcut learning
3. 3Develop strategies to mitigate object-driven shortcuts, such as using temporal evidence instead of relying on labeled object classes
4. 4Implement and evaluate these strategies in zero-shot compositional action recognition models

Machine learning researchers and engineers working on action recognition tasks can benefit from this research to improve their models' accuracy and robustness

💡 Object-driven shortcuts can hinder model performance in zero-shot compositional action recognition, and addressing sparse compositional supervision and verb-object learning asymmetry can help

🤖 Mitigating object-driven shortcuts in zero-shot action recognition #AI #ML