Reproduce “On the Efficiency of Training Robust Decision Trees” 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

Benedict Gerlach, Marie Anastacio, Holger H. Hoos

Researchers empirically analyzed the efficiency of training robust decision tree ensembles, revealing that robustness verification time can dominate the overall computational cost and often shows no correlation with training time. The study found that smaller Random Forest models can efficiently estimate perturbation sizes, and while some robust training methods are fast, their verification can be orders of magnitude slower, impacting overall pipeline efficiency.

Key insights

  • Robustness verification time consistently exhibits exponential or high polynomial growth with increasing tree depth and number of trees, often dominating the total computational cost.
  • There is no direct correlation between a model's training time and its subsequent robustness verification time, meaning fast-training methods can be slow to verify and vice-versa.
  • Significant efficiency gains (factors of 10 to over 1000) can be achieved by estimating perturbation sizes using smaller Random Forest models, especially for hard-to-verify datasets.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.