Skip to content

Latest commit

 

History

History
executable file
·
38 lines (26 loc) · 3.02 KB

README.md

File metadata and controls

executable file
·
38 lines (26 loc) · 3.02 KB
Raw

normalized online learning demo

These demos show the ability of the normalized learning rule to adapt to varying feature scales, relative to an unnormalized learning rule. For more details on the normalized learning rule see the paper.

Instructions

  • make all.results: eventually produces a nice table outlining performance of normalized adaptive gradient (NAG) vs. unnormalized adaptive gradient (AG) for a variety of data sets.
    • WARNING: Please be aware that these demos can be network I/O, disk space, and/or CPU intensive.
    • The complete set of demos can take hours to compute the first time, although the results are cached for subsequent reproduction.
    • You will see lower progressive loss (regret) for NAG than AG.
    • You will also see that the optimal learning rate eta* varies less across datasets for NAG than AG.

Details

This is organized into individual demos which process a single dataset. The different individual demos are associated with unique make targets.

An individual demo will download a data set and learn a predictor under two conditions: using the normalized learning rule aka NAG (this the vw default), and using adaptive gradient without normalization aka AG (invoked via vw arguments --adaptive --invariant). For both conditions it will do a hyper-parameter sweep to find the optimal in-hindsight learning rate eta*. Note the context here is online learning, so there is no train/test split; rather what is optimized is progressive loss over the input data.

About the datasets

Note when data is pre-normalized or otherwise does not exhibit varying scales, the normalized learning rule has essentially no effect. Therefore the data sets used in this demo have been selected because they exhibit varying scales. The data set covertype exemplifies how this arises in practice, as it consists of multiple physical measurements with different units.

Just doing one dataset

There are individual makefile targets that will just download and compare one dataset. You can compute a subset using a combination of make targets, e.g., make only.{covertype,shuttle}.

Invocation Dataset Time Disk Network
make only.bank bank 2 minutes 1Mb 1Mb
make only.census census 10 minutes 13Mb 7Mb
make only.covertype covertype 12 minutes 24Mb 11Mb
make only.CTslice CT Slice 30 minutes 40Mb 18Mb
make only.MSD MSD 30 minutes 500Mb 256Mb
make only.shuttle Shuttle 1 minute 600Kb 300Kb