Models

PPL Bench currently has the following models.

Bayesian Logistic Regression

  • Simple model; baseline
  • Log-concave posterior, easy convergence
  • For a detailed description of this model, go here.

Robust Regression

  • Increased robustness to outliers
  • Uses a Bayesian regression model with Student-T errors
  • For a detailed description of this model, go here.

Noisy-Or Topic Model

  • Inferring topics from words in a document
  • Bayesian Netrowk structure with topics and words as nodes
  • Supports hierarchical topics
  • For a detailed description of this model, go here.

Crowdsourced Annotation

  • Inferring true label of an object given multiple labeler's label assignments
  • Maintain confusion matrix of each labeler
  • Includes inferring the unknown prevalence of labels
  • For a detailed description of this model, go here.

Adding New Models

PPL Bench supports adding new models. Please refer to CONTRIBUTING.md for details on how to do so!