Future Work

Mick Thomure edited this page Jun 28, 2013 · 1 revision
Clone this wiki locally

The following list gives some ideas for improving Glimpse in the future. The list is in no particular order.

  • Integration into a general machine learning framework. Ideally, this framework would provide a graphical interface for designing and running experiments. A good candidate for such a framework is the Orange project.
  • More advanced backends, using:
    • GPUs: Push evaluation of layer-wise operations to a GPU. This could probably be written using PyCUDA or Theano.
    • Vector Intrinsics (SSE): Evaluate a layer-wise operation on several units in parallel. Some code for this exists in old versions of the project, and should be dusted off.
  • A graphical user interface (GUI) that allows the user to specify arbitrary network topologies. This might be done by hacking an interface out of the Orange project's workbench code.
  • App package for OS X, probably using PyInstaller or py2app.
  • Integrated GUI for running experiments and analyzing results. As an example, this should integrate the plots shown in the user guide. A start in this direction has been made using PySide.
  • Automated loader/downloader for image corpora, similar to the mechanism provided by scikit-learn. For example, this should allow the user to download and unpack the AnimalDB dataset with a single command.
  • Add a script to perform classification on many sub-windows of the same image. Use the optimization we built for George's thesis.
  • Add an iterable interface for joblib.Parallel, with access to results as they arrive. This is necessary to support a progress meter for the MulticorePool.