Fergal, I can't follow all of your arguments but I can address one item. There are no differences in the number of segments/synapses/memory required between the old and new method. In both cases a cell has to learn to respond to multiple patterns. In both the old and new TP methods to pool over X patterns requires the same amount of memory and the same amount of training time. The difference is that in the old TP the patterns pooled were states of the local sequence memory and in the new TP method the patterns pooled are input bits coming from a different CLA sequence memory.
Neither case requires "tagging" a segment. Even though learning requires the patterns occur sequentially in time, after learning the cell will respond regardless of the order. This isn't a problem because the states we pool are always high order states of a sequence memory. Jeff From: nupic [mailto:[email protected]] On Behalf Of Fergal Byrne Sent: Tuesday, January 28, 2014 9:27 AM To: NuPIC general mailing list. Subject: Re: [nupic-discuss] New Temporal Pooler Proposal HI Subutai, I think I did see that, but I don't see how it solves the problem in the way that Jeff's new idea does. This is for a number of reasons. First, as you say, the solution is meta-neural (ie you have to come up and have data structures remembering lots of previous states); secondly, you have to do this again for each extension of the sequence into the past, so you would quickly run out of segments; finally (and most importantly), you would lose or dilute the current TP functionality of confidently predicting the next step, because you wouldn't (without tagging each segment) be able to distinguish which timestep(s) made which contribution to the predictive state. Jeff's solution (with my adjustments) provides you with a weighted predictive state in Layer 4 which grows in confidence as more and more elements of the sequence appear, is stable over any length of sequence, shows you roughly where you are in the sequence, and also give you an indication of how to move though the sequence (either in time with rewind, pause and play, or via actual motor commands). It works with the current CLA in Layer 3 to give you both fast and slow-changing SDR's, and also provides a mechanism for Layer 5 to emit motor commands in the context of the sequence (by projecting representations of [target-state, motor command]. On the point of illustrations, absolutely a great idea. Regards, Fergal Byrne On Tue, Jan 28, 2014 at 5:02 PM, Matthew Taylor <[email protected]> wrote: On Tue, Jan 28, 2014 at 8:46 AM, Subutai Ahmad <[email protected]> wrote: One thing we'll want to do with the new proposal is to create detailed step by step illustrations of every aspect. This was sorely missing I think in the whitepaper. It would be awesome if anyone wants to contribute with this. *cough* Patrick Higgins! *cough* --------- Matt Taylor OS Community Flag-Bearer Numenta _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org -- Fergal Byrne, Brenter IT http://inbits.com - Better Living through Thoughtful Technology e:[email protected] t:+353 83 4214179 Formerly of Adnet [email protected] http://www.adnet.ie <http://www.adnet.ie/>
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
