Google Summer of Code 2015 projects

A list of potential GSoC 2015 projects.

To add, please create a new wiki page for each project that you describe. Name them as "GSoC_2015_project_MCMC" etc. Here is a template.

Main focus

This year's GSoC is about improving Shogun, rather than extending it. Exceptions allowed.

• Fewer new algorithms. Rather improve existing ones: Usability, efficiency, documentation, application
• Fewer students. More intense mentoring, interaction between students, blogging, documenting
• Projects on
  • Installation: most important
  • Clean ups of: framework, build process, algorithms, usability, documentation
  • Removing legacy code
  • Efficiency: High performance computing, parallelisation, cloud, benchmarking
  • Applications: Using Machine Learning as a tool to improve the world, rather than toy examples
  • Pipelines & Framework: Improve usability on standard workflow pipelines
  • Usablity: Building and Installing Shogun has to be easier

Mentors (and students) this year

Projects

ALL students will be required to document existing Shogun code on a weekly basis during GSoC. We need a list of all methods in Shogun, along with slim documentation on how to use them. Details to follow!

Improving Shogun

These are roughly ordered in our priority in them. Most of them do not focus on Machine Learning but rather on software engineering.

• Easy installation on major platforms
• Native MS Windows port
• Framework for plugin-based architecture
• A Shogun Detox
• Unifying Shogun's linear algebra
• HMM cleanup and application
• Shogun cloud extensions

Extending Shogun:

The projects we would like to limit in numbers.

Algorithms

• Fundamental ML 2: LGSSMs
• Large Scale Gaussian Processes
• Hip Deep learning
• Solver for the KKT System
• Dual coordinate ascent solver for SO-SVM
• LP/QP Framework
• Density Estimation in Infinite Dimensional Exponential Families
• Debiasing & Cluster computing

Framework

• Independent jobs Framework
• MCMC & Stan

TODO list

Other ideas:

• Cool pipelines
  • A kaggle pipeline for supervised prediction.
  • Spectrometer (there is an open-source hardward project on this)
  • Music brainz predictions (The cool hair guy at GSoC is the one we should talk to here)
  • Some bio thing?
  • Collaboration with MLPack for toolkit wide performance/accuracy testing

Infrastructure:

• Easy Model selection!
• Get rid of static interfaces, migrate all tests etc
• Matlab swig bindings
• REST interface

Clean ups:

• Modularise Shogun
• Replace parameter framework
• Benchmark existing algos and improve!