Using probabilistic programs as proposals. Cusumano-Towner, M. F. & Mansinghka, V. K. In Workshop on Probabilistic Programming Languages, Semantics, and Systems (co-located with POPL 2018), 2018. Paper Link abstract bibtex Monte Carlo inference has asymptotic guarantees, but can be slow when using generic proposals. Handcrafted proposals that rely on user knowledge about the posterior distribution can be efficient, but are difficult to derive and implement. This paper proposes to let users express their posterior knowledge in the form of proposal programs, which are samplers written in probabilistic programming languages. One strategy for writing good proposal programs is to combine domain-specific heuristic algorithms with neural network models. The heuristics identify high probability regions, and the neural networks model the posterior uncertainty around the outputs of the algorithm. Proposal programs can be used as proposal distributions in importance sampling and Metropolis-Hastings samplers without sacrificing asymptotic consistency, and can be optimized offline using inference compilation. Support for optimizing and using proposal programs is easily implemented in a sampling-based probabilistic programming runtime. The paper illustrates the proposed technique with a proposal program that combines RANSAC and neural networks to accelerate inference in a Bayesian linear regression with outliers model.
@inproceedings{towner2018proposals,
title = {Using probabilistic programs as proposals},
author = {Cusumano-Towner, Marco F. and Mansinghka, Vikash K.},
booktitle = {Workshop on Probabilistic Programming Languages, Semantics, and Systems (co-located with POPL 2018)},
year = 2018,
url_paper = {https://arxiv.org/pdf/1801.03612.pdf},
url_link = {https://popl18.sigplan.org/event/pps-2018-using-probabilistic-programs-as-proposals},
keywords = {probabilistic programming, amortized inference, randomized algorithms, neural networks, proposal distribution, importance sampling, metropolis-hastings, auxiliary variable},
abstract = {Monte Carlo inference has asymptotic guarantees, but can be slow when using generic proposals. Handcrafted proposals that rely on user knowledge about the posterior distribution can be efficient, but are difficult to derive and implement. This paper proposes to let users express their posterior knowledge in the form of proposal programs, which are samplers written in probabilistic programming languages. One strategy for writing good proposal programs is to combine domain-specific heuristic algorithms with neural network models. The heuristics identify high probability regions, and the neural networks model the posterior uncertainty around the outputs of the algorithm. Proposal programs can be used as proposal distributions in importance sampling and Metropolis-Hastings samplers without sacrificing asymptotic consistency, and can be optimized offline using inference compilation. Support for optimizing and using proposal programs is easily implemented in a sampling-based probabilistic programming runtime. The paper illustrates the proposed technique with a proposal program that combines RANSAC and neural networks to accelerate inference in a Bayesian linear regression with outliers model.},
}
Downloads: 0
{"_id":"6RNA6PvsRrj7SX7pM","bibbaseid":"cusumanotowner-mansinghka-usingprobabilisticprogramsasproposals-2018","downloads":5,"creationDate":"2018-01-18T16:31:30.698Z","title":"Using probabilistic programs as proposals","author_short":["Cusumano-Towner, M. F.","Mansinghka, V. K."],"year":2018,"bibtype":"inproceedings","biburl":"probcomp.csail.mit.edu/papers.bib","bibdata":{"bibtype":"inproceedings","type":"inproceedings","title":"Using probabilistic programs as proposals","author":[{"propositions":[],"lastnames":["Cusumano-Towner"],"firstnames":["Marco","F."],"suffixes":[]},{"propositions":[],"lastnames":["Mansinghka"],"firstnames":["Vikash","K."],"suffixes":[]}],"booktitle":"Workshop on Probabilistic Programming Languages, Semantics, and Systems (co-located with POPL 2018)","year":"2018","url_paper":"https://arxiv.org/pdf/1801.03612.pdf","url_link":"https://popl18.sigplan.org/event/pps-2018-using-probabilistic-programs-as-proposals","keywords":"probabilistic programming, amortized inference, randomized algorithms, neural networks, proposal distribution, importance sampling, metropolis-hastings, auxiliary variable","abstract":"Monte Carlo inference has asymptotic guarantees, but can be slow when using generic proposals. Handcrafted proposals that rely on user knowledge about the posterior distribution can be efficient, but are difficult to derive and implement. This paper proposes to let users express their posterior knowledge in the form of proposal programs, which are samplers written in probabilistic programming languages. One strategy for writing good proposal programs is to combine domain-specific heuristic algorithms with neural network models. The heuristics identify high probability regions, and the neural networks model the posterior uncertainty around the outputs of the algorithm. Proposal programs can be used as proposal distributions in importance sampling and Metropolis-Hastings samplers without sacrificing asymptotic consistency, and can be optimized offline using inference compilation. Support for optimizing and using proposal programs is easily implemented in a sampling-based probabilistic programming runtime. The paper illustrates the proposed technique with a proposal program that combines RANSAC and neural networks to accelerate inference in a Bayesian linear regression with outliers model.","bibtex":"@inproceedings{towner2018proposals,\ntitle = {Using probabilistic programs as proposals},\nauthor = {Cusumano-Towner, Marco F. and Mansinghka, Vikash K.},\nbooktitle = {Workshop on Probabilistic Programming Languages, Semantics, and Systems (co-located with POPL 2018)},\nyear = 2018,\nurl_paper = {https://arxiv.org/pdf/1801.03612.pdf},\nurl_link = {https://popl18.sigplan.org/event/pps-2018-using-probabilistic-programs-as-proposals},\nkeywords = {probabilistic programming, amortized inference, randomized algorithms, neural networks, proposal distribution, importance sampling, metropolis-hastings, auxiliary variable},\nabstract = {Monte Carlo inference has asymptotic guarantees, but can be slow when using generic proposals. Handcrafted proposals that rely on user knowledge about the posterior distribution can be efficient, but are difficult to derive and implement. This paper proposes to let users express their posterior knowledge in the form of proposal programs, which are samplers written in probabilistic programming languages. One strategy for writing good proposal programs is to combine domain-specific heuristic algorithms with neural network models. The heuristics identify high probability regions, and the neural networks model the posterior uncertainty around the outputs of the algorithm. Proposal programs can be used as proposal distributions in importance sampling and Metropolis-Hastings samplers without sacrificing asymptotic consistency, and can be optimized offline using inference compilation. Support for optimizing and using proposal programs is easily implemented in a sampling-based probabilistic programming runtime. The paper illustrates the proposed technique with a proposal program that combines RANSAC and neural networks to accelerate inference in a Bayesian linear regression with outliers model.},\n}\n\n","author_short":["Cusumano-Towner, M. F.","Mansinghka, V. K."],"key":"towner2018proposals","id":"towner2018proposals","bibbaseid":"cusumanotowner-mansinghka-usingprobabilisticprogramsasproposals-2018","role":"author","urls":{" paper":"https://arxiv.org/pdf/1801.03612.pdf"," link":"https://popl18.sigplan.org/event/pps-2018-using-probabilistic-programs-as-proposals"},"keyword":["probabilistic programming","amortized inference","randomized algorithms","neural networks","proposal distribution","importance sampling","metropolis-hastings","auxiliary variable"],"metadata":{"authorlinks":{"mansinghka, v":"http://probcomp.csail.mit.edu/"}}},"search_terms":["using","probabilistic","programs","proposals","cusumano-towner","mansinghka"],"keywords":["probabilistic programming","amortized inference","randomized algorithms","neural networks","proposal distribution","importance sampling","metropolis-hastings","auxiliary variable"],"authorIDs":["azzu48Y9BFosGy6pT"],"dataSources":["KjwPBikSY3GLxsjaT","PGvySWpKSvwM9TdfF"]}