Output details
11 - Computer Science and Informatics
University of Edinburgh
A model-learner pattern for bayesian reasoning
<08> Originality: Identifies a probabilistic programming abstraction for a wide variety of Bayesian models in machine learning. It is standard that Bayesian reasoning is based on a prior and likelihood; the paper captures the abstraction in programming language terms for the first time.
Significance: The model-learner pattern allows the generic expression of tasks such as model testing, and generic compositions including mixture models, evidence-based model averaging, and mixtures of experts. This pattern from POPL'13 is the basis of a new programming language Tabular accepted for publication at POPL'14.
Rigour: Formal semantics in measure theory; three separate practical implementations.