Hi All,
This week we will have Philippe Hamel present a python library that he
and Simon have been working on. Hope to see you all there.
Where: LISA lab (AA3256)
When: 14h00, Thursday, July 7th, 2011
Title/Abstract:
Developing tools for hyper-parameter visualization
As the complexity of machine learning algorithms becomes greater, the
number of hyper-parameters tends to do the same. Optimizing such
algorithms often involves tedious exploration of a high dimensional
hyper-parameter space. Typically, results from experiments are stored
in a SQL database and results can be analyzed by looking at sql
queries in a table. With this kind of view, it can be hard to see
correlations between parameters, or to understand trends in the data.
Using plots instead of tables can allow for better understanding of
the hyper-parameter space, but might require more coding. So, a tool
that allows us to easily plot and visualize the hyper-parameter space
could help to guide our experiments in the right direction.
In this tea talk, I would like to present a small python library that
Simon and I have been working on and using to visualize our
experiments. It is basically a bridge between a jobman SQL database
and pylab. It allows dynamic filtering of the data, applying functions
to transform and/or combine columns of data, and more! We have a few
ideas on features that could be added, and we would like to get your
input.
We would like to give you a quick demo of what the library can do for
now. We'd love some feedback from you. We are interested to know if
people from LISA might be interested to use this tool, or even
contribute to it.
Cheers,
Aaron
--
Aaron C. Courville
Département d’Informatique et
de recherche opérationnelle
Université de Montréal
email:Aaron.Courville@gmail.com
