A probabilistic program that writes another program RRS feed

  • Question

  • The Cambridge, MA company called Gamalon have sparked my interest since they apparently use probabilistic programming under the hood. From what I could understand from the examples (can't post the link but the videos are on their site), a top-level probabilistic programming model has been implemented that is flexible enough to "write" probabilistic programs that model the data.

    My question is, is this something that is viable? Using Infer.NET? If not, what kind of technologies/methodologies are missing?

    An example in particular that I am interested in, is automatic text structurization from unstructured text records. The data is many raws of say products that can vary in formatting from data source, provider, time etc. The output is a new table where the information has been broken automatically into columns (e.g. brand, product name, feature, quantity). I have worked through the experimental examples that show how to work with simple distributions over text. I am wondering whether I could use existing Infer.NET functionality to do something similar.

    Wednesday, January 17, 2018 6:10 AM

All replies

  • To solve your example, all you need to do is write a model that generates the unstructured text when given a table.  The inference engine does all the work of reasoning backwards.  You don't need a program that writes another program.
    Wednesday, January 17, 2018 12:00 PM