Hello,

After the warm reception of the Kaggle competition idea, I decided to
create the thread asap. For those who are not aware, Kaggle is a data
mining competition website. It has been hyped on the machine learning media
and has some great people participating in it. It rewards with different
prizes, but my focus is not earn money, but rather to see how good NuPIC
would fare against the state of the art. I found those following
competitions to be a viable start:

1) Accelerometer Biometric Competition
>From real acceleration data you have to guess from which cell phone came
the test samples.
36 days to go
Pros:
-Multivariable, temporal data seems ideal for the CLA
Cons:
-The best competitors are using data leaks (meta-information), so the
chance of winning fairly is zero (though they will give a consolation prize
to the best "honest" entry)
Questions:
-How fast would NuPIC go through a 1GB CSV file?
-The test data is not open-ended, that is, you have to say whether a
recording belongs to a specific cell phone or not, with 50% chance of being
right. Would NuPIC do well in this case?

2) Multi-Label Bird Species Classification
Try to guess the bird species present on some audio recordings
38 days to go
Pros:
-Temporal data
Cons:
-Multiple labels, so I think we would need 87 different models, which is
kinda impractical
Question:
-Is there a better way to do multi-label classification with NuPIC?

3) Conway's Reverse Game of Life
Reverse the game of life for 1 to 5 steps
4 months to go
Pros:
-Natural input sparseness (have to check on it)
-Temporal data
-Multistep classification
-We can generate our own training data if necessary
Cons:
-Maybe there are very straightforward ways to solve this (rule based, brute
force, etc)
-Too much time before the competition ends
Questions:
-How sparse does the input need to be for the CLA to work well?
-Is it possible to use a 2D arrangement on the cells connection matrix?

There is also the AMS 2013-2014 Solar Energy Prediction contest, but I
didn't have time to go over it. The other competitions don't seem to be
suited for NuPIC.

Who here is interested on participating and what competition do you think
to be the most promising one?

Pedro.
-- 
Pedro Tabacof,
Unicamp - Eng. de Computação 08.
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to