Stochastic Average Gradient

Background

Mark Schmidt proposed the Stochastic Average Gradient (SAG) algorithm as a fast solver for smooth convex optimization problems on finite data sets. His C/MATLAB code implements SAG for L2-regularized logistic regression. L2-regularized logistic regression is a convex optimization problem that is explained in detail in Chapter 4 of Elements of Statistical Learning.

Related work

• glmnet(family="binomial", alpha=0) uses coordinate descent to solve L2-regularized logistic regression.
• optimx is a general function optimizer which could be used to solve L2-regularized logistic regression.

Project ideas

Fork https://github.com/tdhock/SAG and write the SAG R package:

• Convert Mark's C code with "mex.h" headers to C code with "R.h" headers, for the three SAG methods (SAG, SAGlineSearch, SAG_LipshitzLS).
• Convert Mark's documentation comments in C code to .Rd files, possibly generated by inlinedocs, etc.
• Examples/vignettes using Mark's rcv1_train and covtype.libsvm data sets that
  • show how these 3 solvers can be used,
  • compare with the results of glmnet/optimx.
• Tests that make sure the R package
  • gets the right answer (gradient with norm close to zero).
  • gets the same answer as glmnet/optimx.

Skills required

R package and C code development.

Mentor

Please get in touch with John Nash [email protected] and Toby Dylan Hocking [email protected] as soon as possible.

Tests

After completing your tests, please post a link to your files below.

• Easy: use glmnet to fit an L2-regularized logistic regression model. Use the system.time function to record how much time it takes for several data set sizes, and make a plot that shows how execution time depends on the data set size.
• Medium: create a simple R package with one function and one documentation file, and upload it to your GitHub account.
• Harder: Write an R package which uses .C to interface C code.