Ablation Study
A systematic experiment where components of a model or system are removed one at a time to measure the contribution of each component to overall performance.
Purpose
Ablation studies answer 'which parts of this system actually matter?' By removing components (layers, features, training techniques) and measuring the impact, researchers identify what drives performance and what is unnecessary.
In Practice
A typical ablation might remove attention heads, skip connections, data augmentation, or specific loss terms. The performance difference reveals the contribution of each component.
Importance in Research
Top AI conferences require ablation studies in paper submissions. They prevent over-engineering by showing which components are essential and which can be simplified.