Reproduce “A Tunable Loss Function for Robust Classification: Calibration, Landscape, and Generalization” 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

Tyler Sypherd, Mario Diaz, John Kevin Cava, Gautam Dasarathy, Peter Kairouz, Lalitha Sankar

A new tunable loss function, α-loss, is introduced and analyzed for classification tasks, unifying exponential and log-losses while demonstrating strong information-theoretic foundations. It empirically improves model robustness against noisy labels when α > 1 and increases sensitivity to minority classes under imbalance when α < 1, often surpassing log-loss performance.

Key insights

  • α-loss unifies common loss functions like exponential, log, and 0-1 loss, providing a continuous framework for understanding and comparing their properties.
  • The parameter α can be tuned to adapt the learning objective: α > 1 enhances robustness against noisy labels, while α < 1 increases sensitivity to minority classes in imbalanced datasets.
  • The theoretical analysis reveals how α influences the loss function's convexity, quasi-convexity, and overall optimization landscape, including a saturation effect for large α.

Summary by alphaXiv.

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