Thanks for the correction, Felix.

---------
Matt Taylor
OS Community Flag-Bearer
Numenta

On Tue, Apr 12, 2016 at 6:22 PM, Felix Andrews <[email protected]> wrote:

> There is overfitting in the sense of learning overly specific patterns
> without generalising. But that's a much broader question than just
> swarming, and is unsolved.
>
> -f
>
>
>
> On 13 April 2016 at 05:59, Matthew Taylor <[email protected]> wrote:
>
>> Hi Sam. No danger of overfitting with HTM because it is an online
>> learning system. It changes it's representation of the input space as the
>> temporal patterns change. Don't think of swarming as a "test set", just
>> think of it as a data sample to get the right encoding parameters.
>>
>> ---------
>> Matt Taylor
>> OS Community Flag-Bearer
>> Numenta
>>
>> On Tue, Apr 12, 2016 at 2:20 PM, Samuel O Heiserman <
>> [email protected]> wrote:
>>
>>> Hey Nupic!
>>>
>>>     I'm wondering: when running data ]through Nupic, should I not run
>>> the same file to build the model as I did to swarm for the parameters?
>>> Since the parameters were tuned to that exact data, it seems like a
>>> potential overfitting risk. The data is a series of control actions of
>>> subjects playing a simple game. What I'm trying to do is train a model on
>>> the subject 1's data, save that model and use it to forecast for subjects 1
>>> - 20.
>>>     I hope to show that the HTM can learn the individual behavioral
>>> patterns of a given subject distinct from the others, and I plan to show
>>> this capacity with a result where the model does well forecasting for all
>>> subjects, but especially well at forecasting for the subject it was trained
>>> on. However I wonder if when testing the model on subject 1, I should use
>>> different subject 1 data than I used to swarm for the parameters. Thanks
>>> again!
>>>
>>> -- Sam
>>>
>>
>>
>
>
> --
> Felix Andrews / 安福立
> http://www.neurofractal.org/felix/
>

Reply via email to