context bias

**Context Bias** is the **reliance of models on co-occurring objects, scene context, or spatial relationships for classification** — the model learns that certain objects always appear together (e.g., keyboard with monitor) and uses context cues rather than object-specific features for prediction. **Context Bias Examples** - **Co-Occurrence**: "Tennis racket" prediction relies on detecting "tennis court" or "tennis ball" in the image. - **Spatial Context**: Object detection accuracy depends on where in the scene the object appears — unusual positions cause misses. - **Scene Priors**: Indoor scenes bias toward "furniture" classes, outdoor toward "vehicles" — regardless of actual content. - **Language Bias**: In VQA, models learn statistical priors ("What color is the banana?" → "yellow") without looking at the image. **Why It Matters** - **Counter-Intuitive Scenes**: Models fail on unusual contexts — a boat on land, a car in a living room. - **Out-of-Context Detection**: Context bias undermines the ability to detect objects in unusual settings. - **Causal vs. Correlational**: Models learn correlational context rather than causal features of the target object. **Context Bias** is **guilt by association** — classifying objects based on their usual companions rather than their own distinctive features.