Question about input values

[Erik Chelstad]

Hi, everyone,

I was just wondering if anyone has experience using a binary representation
of an input value vs. the float.

For instance if an input varies between say 0 and 500,000, would the network
come to a more optimal with one float value scaled down to 0-1, or 19
individual inputs taking a binary number?

I have not done the experiment yet, but was wondering perhaps if anyone else
had, or if there was a paper that talks about it.

Thanks,

Erik

[Josh Menke]

Hi Erik,

If order matters (like, 5 really is > 4) then I would standardize the number
to mean 0 and variance 1 (subtract the mean across all records, and divide
by the standard deviation).

More inputs means more parameters (weights) which means it will take longer
to train, and is harder to keep from overfitting.

I would definitely go with standardizing. Even better would be to reduce
covariance with other inputs.

--Josh

Post by Erik Chelstad
Hi, everyone,
I was just wondering if anyone has experience using a binary
representation of an input value vs. the float.
For instance if an input varies between say 0 and 500,000, would the
network come to a more optimal with one float value scaled down to 0-1, or
19 individual inputs taking a binary number?
I have not done the experiment yet, but was wondering perhaps if anyone
else had, or if there was a paper that talks about it.
Thanks,
Erik
-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share
your
opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV
_______________________________________________
Fann-general mailing list
https://lists.sourceforge.net/lists/listinfo/fann-general

[Erik Chelstad]

Thanks, Josh.

Your comments made me take a different route than I was originally thinking,
and I used an approach that was more like a variation.

Can't say how the results are just yet, as I'm playing around with various
GA libraries to test out a variety of inputs.

- Erik

Post by Josh Menke
Hi Erik,
If order matters (like, 5 really is > 4) then I would standardize the
number to mean 0 and variance 1 (subtract the mean across all records, and
divide by the standard deviation).
More inputs means more parameters (weights) which means it will take
longer to train, and is harder to keep from overfitting.
I would definitely go with standardizing. Even better would be to reduce
covariance with other inputs.
--Josh

Post by Erik Chelstad
Hi, everyone,
I was just wondering if anyone has experience using a binary
representation of an input value vs. the float.
For instance if an input varies between say 0 and 500,000, would the
network come to a more optimal with one float value scaled down to 0-1, or
19 individual inputs taking a binary number?
I have not done the experiment yet, but was wondering perhaps if anyone
else had, or if there was a paper that talks about it.
Thanks,
Erik
-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share
your
opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV
_______________________________________________
Fann-general mailing list
https://lists.sourceforge.net/lists/listinfo/fann-general

--
Joshua Menke
Machine Learning Scientist
Trust and Safety Applied Research
ebay, Inc
-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share
your
opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV
_______________________________________________
Fann-general mailing list
https://lists.sourceforge.net/lists/listinfo/fann-general