David (Wood), don't worry, this particular topic still confuses lots of
people ;)

David (cogmission), you configure it using the MultiStepPrediction setting,
giving a list of numbers indicating how far ahead to predict (= how far
behind to record).

On Thu, Mar 5, 2015 at 11:07 PM, David Wood <[email protected]> wrote:

> Thanks, Chandan and David. That’s helpful. Sorry for the newbie question!
>
> Regards,
> Dave
> --
> http://about.me/david_wood
>
>
>
> On Mar 5, 2015, at 17:17, cogmission <[email protected]> wrote:
>
> David,
>
> The CLAClassifier is configured to compute the probability of all the
> buckets at the specified steps. I'm not sure how one goes about configuring
> this from the OPF or NetworkAPI point of view. The TemporalMemory (new
> sequence memory) always predicts the next step (at t + 1). You can pull out
> the prediction for the step in question by querying the CLAClassifier or if
> the step == 1, you can simply use the results of the TemporalMemory's last
> cycle to get the predicted column SDR for the next cycle.
>
> David
>
> On Thu, Mar 5, 2015 at 4:09 PM, David Wood <[email protected]> wrote:
>
>> Hi Fergal,
>>
>> I’m a bit confused by your answer. Is there no way using NuPIC to
>> estimate multi-step predictions?
>>
>> Regards,
>> Dave
>> --
>> http://about.me/david_wood
>>
>>
>>
>> On Mar 5, 2015, at 15:49, Fergal Byrne <[email protected]>
>> wrote:
>>
>> Hi Michael,
>>
>> The region itself always just predicts one step ahead. You can connect a
>> region with code (most of it in OPF) which will remember what happens N
>> steps ahead of a timestep, but this is just a histogram record (associating
>> a cell's activation with an input field value) of what is likely to come up
>> after N steps. This is what is used if you specify multi-step predictions.
>>
>> Ignore the multi-step stuff in the White Paper. It's wrong, and has been
>> abandoned. CLA on its own just does a single timestep prediction, and this
>> is what also happens in neocortex.
>>
>> Regards,
>>
>> Fergal Byrne
>>
>>
>> On Thu, Mar 5, 2015 at 12:38 AM, cogmission <[email protected]>
>> wrote:
>>
>>> Oh the Prediction code is in CLAClassifier and the Anomaly code does the
>>> running total of the meta qualities...
>>>
>>> On Wed, Mar 4, 2015 at 6:36 PM, cogmission <[email protected]>
>>> wrote:
>>>
>>>> Hi Michael,
>>>>
>>>> Afaik, the "Anomaly" class is what you are looking for, just that it
>>>> tracks the moving average of accuracy or maybe the inverse (anomaly). You
>>>> could in any case have a look at that code to see if it either does what
>>>> you are looking for or can be "adapted" to do more of what you're looking
>>>> for.
>>>>
>>>> Also afaik, the steps will "overwrite" when that point in the cycle is
>>>> reached again (so every 500 steps a new prediction quality is estimated -
>>>> if 500-steps is one of the step configurations).
>>>>
>>>> Correct me if I'm wrong someone?
>>>>
>>>> David
>>>>
>>>> On Wed, Mar 4, 2015 at 6:21 PM, Michael Roy Ames via nupic <
>>>> [email protected]> wrote:
>>>>
>>>>>
>>>>>
>>>>> ---------- Forwarded message ----------
>>>>> From: Michael Roy Ames <[email protected]>
>>>>> To: NuPIC Mailing List <[email protected]>
>>>>> Cc:
>>>>> Date: Wed, 04 Mar 2015 16:08:38 -0800
>>>>> Subject: Prediction. Several steps. Future or past.
>>>>>  NuPIC list:
>>>>>
>>>>> "Predictions in an HTM region can be for several time steps into the
>>>>> future" - according to the HTM White paper.
>>>>>
>>>>> Question 1: Is there a NuPIC code that does prediction for the next n
>>>>> time steps?
>>>>>
>>>>> Question 2: Is there NuPIC code that keeps activation history such
>>>>> that one could access the last 15 or 20 sets of active cells?
>>>>>
>>>>> I'm interested in making NuPIC learn and recognize temporal sequences
>>>>> of data, and want to limit the amount of additional code I have to write 
>>>>> to
>>>>> get this done. So, I'd rather use existing NuPIC functionality that works
>>>>> instead of writing algorithm that might duplicate something already in
>>>>> place. The sequences may be long (500 steps) or short (20 steps). The
>>>>> one-step predictions I've found in NuPIC examples need extra code to be
>>>>> written to 'remember' the predictions and how many predictions in-a-row
>>>>> have been correct, each additional successful prediction lending greater
>>>>> confidence to the data recognition.
>>>>>
>>>>> Question 3: Is there code that does this already (successful
>>>>> prediction tracking), or will I have to write it?
>>>>>
>>>>> MRA
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> *We find it hard to hear what another is saying because of how loudly
>>>> "who one is", speaks...*
>>>>
>>>
>>>
>>>
>>> --
>>> *We find it hard to hear what another is saying because of how loudly
>>> "who one is", speaks...*
>>>
>>
>>
>>
>> --
>>
>> Fergal Byrne, Brenter IT
>>
>> http://inbits.com - Better Living through Thoughtful Technology
>> http://ie.linkedin.com/in/fergbyrne/ - https://github.com/fergalbyrne
>>
>> Founder of Clortex: HTM in Clojure -
>> https://github.com/nupic-community/clortex
>>
>> Author, Real Machine Intelligence with Clortex and NuPIC
>> Read for free or buy the book at https://leanpub.com/realsmartmachines
>>
>> Speaking on Clortex and HTM/CLA at euroClojure Krakow, June 2014:
>> http://euroclojure.com/2014/
>> and at LambdaJam Chicago, July 2014: http://www.lambdajam.com
>>
>> e:[email protected] t:+353 83 4214179
>> Join the quest for Machine Intelligence at http://numenta.org
>> Formerly of Adnet [email protected] http://www.adnet.ie
>>
>>
>>
>
>
> --
> *We find it hard to hear what another is saying because of how loudly "who
> one is", speaks...*
>
>
>


-- 

Fergal Byrne, Brenter IT

http://inbits.com - Better Living through Thoughtful Technology
http://ie.linkedin.com/in/fergbyrne/ - https://github.com/fergalbyrne

Founder of Clortex: HTM in Clojure -
https://github.com/nupic-community/clortex

Author, Real Machine Intelligence with Clortex and NuPIC
Read for free or buy the book at https://leanpub.com/realsmartmachines

Speaking on Clortex and HTM/CLA at euroClojure Krakow, June 2014:
http://euroclojure.com/2014/
and at LambdaJam Chicago, July 2014: http://www.lambdajam.com

e:[email protected] t:+353 83 4214179
Join the quest for Machine Intelligence at http://numenta.org
Formerly of Adnet [email protected] http://www.adnet.ie

Reply via email to