Ablation studies is quite common in machine learning literature, especially those on neural network and deep learning. When a new architecture or method is proposed, the authors often perform an ablation study where they remove sub-components of their methods that they feel are important one at a time, and study how the performance of said method changes, to learn the actual importance of each of the sub-components.

I find this approach quite useful for research on optimization algorithms, specifically heuristical algorithms, whose properties may not be easy or even possible to prove mathematically.

However, I have only seen ablation study or the term "ablation study" (might be called something else in other fields) in machine learning research. Would it be appropriate to include in papers on mathematical optimization algorithms or other areas of research in general?

More Tianxun Zhou's questions See All
Similar questions and discussions