GSoC_2015_project_dca_sosvm

Heiko Strathmann edited this page Feb 20, 2015 · 1 revision
Clone this wiki locally

A dual coordinate ascent solver for SO-SVM

Mentors

Difficulty & Requirements

Medium.

You need to be able to

  • very good in C/C++
  • basics of optimization

Description

The structured output Support Vector Machines (SO-SVM) is a supervised method for learning parameters of a linear classifier with possibly huge number of classes. Learning is translated into a convex optimization program size of which scales with the number of classes making the problem intractable by off-the-shelf solvers. A dual coordinate ascent (DCA) based methods, well known e.g. from two-class SVMs, have been recently proposed for solving the SO-SVM [1][2]. The DCA algorithms combine advantages of approximate online solvers and precise cutting plane methods used so far. In particular, the DCA algorithms process training example in a on-line fashion by a simple update rule and they provide a certificate of optimality at the same time. The goal of this project will be to implement several variants of the DCA algorithms published in [1][2], namel, the Block-coordinate Frank-Wolfe and FASOLE.

[1] S.Lacoste-Julien, M.Jaggi, M.Schmidt, and P.Pletscher. Block-coordinate Frank-Wolfe optimization for structural SVMs. ICML, 2013. [2] V.Franc: FASOLE. Fast Algorithm for Structured Output LEarning. ECML, 2014.

Waypoints and initial work

  • Reading the papars [1][2] to get familiar with BCFW and FASOLE.
  • Getting familiar with Shogun's framework for SO learning.
  • Implementing BCFW.
  • Implementing FASOLE.
  • Benchmarking, comparison with methods already in Shogun.
  • Writing documentation.

Why this is cool

The project is about implementing the current state-of-the-art solvers for SO learning which has many exciting application in domains like computer vision, bio-informatics, text processing and so on.

Useful ressources