Unsupervised program synthesis

The general idea is to find the most compressive representation of the data, where we consider representations that are of the form f(x_i), where f is a program to be synthesized and x_i is an unobserved argument to that program. Through some noise model, f(x_i) produces the ith observation.

Thus, we compress, or "squash" the data, hence the name Sasquatch.

To run this software on the SVRT problems:

• setup a python virtual environment
• install Z3 locally into the virtual environment
• clone this repo into the virtual environment
• Download SVRT to gain access to the images
• run python parse.py <n> where <n> is an SVRT problem number

Dependencies

• Z3

Project Structure

The project includes the following files:

• README: this file
• setup.py: installation information for pip
• bin/: the executable files associated with this project
  • experiment.sh: a wrapper for running iterative experiments on MIT's athena computers
• sasquatch/: the project's python files
  • sasquatch.py: the core wrapper around Z3 for squashing via program synthesis
  • experiments/: applications of Sasquatch
    • morphology/: an experiment in discovering programs for English morphology
      • celex.py
      • corpus.py
      • ipa.py
      • language.py
      • lexicon.py
      • loop_language.py
      • morphology.py
      • morphology_baseline.py
      • process_verbs.py
      • verbs
    • regress/: an experiment in discovering programs for linear regression
      • regress.py
    • svrt/: an experiment in discovering programs for SVRT image classification
      • classifier.py
      • dinnerParty.py: an implementation of parallel mapping
      • find_bad_parses.py
      • parse.py
      • parser_utilities.py
      • vision.py
      • vision_data.py
      • vision_notes
      • analysis/:
        • accuracies.py
        • curves.py
        • figures.py
      • final_data/:
        • many data files
      • neural/:
        • collect_data.py
        • neural.lua