This is the Splice Site recognition dataset from the 2008 Pascal Large
Scale Learning Challenge (http://largescale.ml.tu-berlin.de/summary/).

=== WARNINGS ===

  * you need a beefy machine to comfortably run these demos 
      4 cores or more
      SSD or lots of RAM because the demo is I/O bound
  * the demos are intolerably slow under cygwin (pipes do not work well)

=== INSTRUCTIONS ===

  * make dna.perf
      downloads the dna dataset, trains a 4-gram logistic regression
      in parallel on 4 cores and computes test set statistics
        disk space requirements: about 3 gigabytes
        training time requirements: about 7 cores for about 15 minutes
          ultimately I/O bound: bzcat is the limiting factor
        memory requirements: less than 512 megabytes
      results in APR of 0.512

  * make dnann.perf
      same as dna.perf, but with additionally 1 neural network hidden node 
      slower (by circa 60 seconds) but better
      results in APR of 0.532

  * make dnasmash.perf
      as above but builds a better model
      uses 10 iteratons of parallel sgd with 4 neural network nodes
        disk space requirements: about 10 gigabytes
          additional 7 gigabytes of vw cache on top of original data
        training time requirements:
          16 minutes to build cache over first (ever) pass
          subsequently, 6 minute per pass if you have SSD or enough RAM cache
            10 passes = 60 minutes (x 6 cores)
      results in APR of 0.545

  * make dnahogwild.perf
      same as dna.perf, but trained via lock-free multicore sgd ("hogwild") 
        rather than parallel sgd + averaging
      nondeterministic, but a typical result is APR of 0.516

  * make dnahogwildnn.perf
      same as dnann.perf, but trained via lock-free multicore sgd ("hogwild")
        rather than parallel sgd + averaging
      nondeterministic, but a typical result is APR of 0.536