Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
Bayesian synthesis of probabilistic programs for automatic data modeling
Saad F., Cusumano-Towner M., Schaechtle U., Rinard M., Mansinghka V. Proceedings of the ACM on Programming Languages3 (POPL):1-32,2019.Type:Article
Date Reviewed: Jun 3 2020

Somewhat overshadowed by neural networks (NN) is another thread in machine learning: the Bayesian-based approach. Less data hungry, it also has the promise of being closer to explainable artificial intelligence (XAI), although it hasn’t had the spectacular successes achieved by NN. This isn’t to say that tremendous advances haven’t also been happening in Bayesian learning.

One distinct advantage of probabilistic programming (PP) is that the tools from modern programming language theory and practice readily apply. The work here is an excellent example of the results of doing just that: by leveraging both a programming language (Venture, in this case) built expressly for PP and the power of embedded domain-specific languages (DSL) for the syntactic representation of models, one can start to infer classes of models instead of just parameters from a given model.

The paper shows, through well-chosen examples, the power of this approach. By choosing the right metalanguage of models, it is possible to infer from data which model feature appears to be present, and in so doing, turning inference into model synthesis. This is an extremely powerful paradigm.

The paper presents its work very rigorously, with Section 4 in the theorem-proof style and section 5 covering full denotational semantics. However, those parts are not an easy read; one needs to be familiar with probability theory, PP, and current programming language theory to understand them in full. The experimental results in Section 7 are the (quite readable) payoff. A nice variety of datasets are analyzed, showing that model inference works well, often. The authors’ method certainly seems promising and powerful.

I would recommend this paper to anyone interested in explainable machine learning. The work presented here is rigorous, yet the results are simply grasped, as are the advantages.

Reviewer:  Jacques Carette Review #: CR146984 (2009-0227)
Bookmark and Share
  Editor Recommended
Featured Reviewer
 
 
Data Models (H.2.1 ... )
 
 
Probabilistic Computation (F.1.2 ... )
 
 
Semantics (D.3.1 ... )
 
 
Probability And Statistics (G.3 )
 
Would you recommend this review?
yes
no
Other reviews under "Data Models": Date
A transient hypergraph-based model for data access
Watters C., Shepherd M. ACM Transactions on Information Systems 8(2): 77-102, 2001. Type: Article
Jun 1 1991
Toward a unified framework for version modeling in engineering databases
Katz R. ACM Computing Surveys 22(4): 375-409, 2001. Type: Article
Feb 1 1993
Graph data model and its data language
Kunii H., Springer-Verlag New York, Inc., New York, NY, 1990. Type: Book (9780387700588)
Dec 1 1991
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy