Hi John,

HTM algorithms are really best suited for streaming, temporal data
sequences. This is how humans operate as well.

You can sometimes contort HTM's to learn a static function but it's trying
to fit a square peg into a round hole. In my experience it doesn't usually
do any better (and often does worse) than simple conventional techniques.

--Subutai

On Wed, Oct 1, 2014 at 10:43 AM, John Blackburn <[email protected]>
wrote:

> Thanks, Subutai. Yes, this does help! One question: traditional neural
> nets can be thought of as a function where you have input and output and
> you train it (supervised learning) to give the right output for given
> input. These NNs are not sequence based. I understand HTM is more advanced
> than this, but can it behave like a function as well? Eg, if tilt(t) =
> F[temp(t)] can HTM find this function? (current value from current value)
> or can it only discover:
>
> tilt(t) = F[tilt(t-1), tilt(t-2),..., temp(t-1), temp(t-2)..]  ?
> (ie current value from past values)
>
>
> On Tue, Sep 30, 2014 at 7:28 PM, Subutai Ahmad <[email protected]>
> wrote:
>
>> On Tue, Sep 30, 2014 at 3:47 AM, John Blackburn <
>> [email protected]> wrote:
>>
>>> What I want it to do is primarily predict tilt(t+1) from temperature(t),
>>> I do not want it to predict temperature(t+1) from temperature(t) as this
>>> will not work well with the data. (unlike Hotgym, the temperature cannot be
>>> predicted from previous temperatures)
>>>
>>
>> John,
>>
>> I'm not sure if this is a terminology thing, but I want to clarify the
>> above statement to make sure we're on the same page. If tilt is the
>> predicted field, NuPIC will always try to predict tilt(t+1) from tilt(t),
>> tilt(t-1), title(t-2), ..., etc. It will only include temperature(t) if it
>> helps. Note that if temperature(t) is completely correlated with tilt(t),
>> adding temperature(t) will not help, because the information is already
>> available in tilt(t).
>>
>> So, the swarm does not look for correlations between fields. It only
>> looks to see which fields add value over and above the past values of the
>> predicted field.
>>
>> My apologies if this was already obvious to you!  This use of past values
>> is a very big difference from traditional batch analytics, so just wanted
>> to be clear.
>>
>> --Subutai
>>
>>
>

Reply via email to