Bootstrapping in a language of thought

Karl DeVries presented ch. 2 (“Bootstrapping in a language of thought: a formal model of conceptual change in number word learning”) of Steve Piantadosi’s 2011 dissertation (Learning and the language of thought) on March 8.

Steve is currently developing LOTlib, a python library to model learning complex concepts as compositions of primitives in a language of thought. He graciously agreed to put together a demo for us on the natural number acquisition model in ch. 2 based on this library.

The main py file is NakedDemo.py. You will also need to have the following three py files in the same folder: BasicPrimitives.py, Miscellaneous.py and Objects.py. A sample output file storing every 1000th Metropolis-Hastings iteration of a fairly long run of this demo is available here: sample-output.xls.

[Update, March 11 2012] Thanks to Robert Henderson, we are aware of a compatibility issue with scipy (which needs to installed to run the demo; numpy needs to be installed too). In the most recent version of scipy, the maxentropy module has been deprecated and logsumexp has been moved to scipy.misc. So if you install a new version of scipy, you must edit Miscellaneous.py and replace “from scipy.maxentropy import logsumexp” with “from scipy.misc import logsumexp”.