Hi Wakan,
For this reason, the term "Sequence Memory" is thought to be more accurate by
some.
The TM does in fact keep track of "what follows what" (and therefore by
extension it knows "what precedes what"). It does this by growing connections
(distal dendrites), to cells which were activated in the preceding time step,
and by rewarding (incrementing synapse permanences) those active cells
("predictive cells") which already have a connection to the current cell.
However, you are correct: the TM does not have a sense of actual time; but it's
nature is "Temporal" because order is a function (attribute of) time.
Cheers,
D
Sent from my iPhone
> On Jan 12, 2016, at 6:28 AM, Wakan Tanka <[email protected]> wrote:
>
> Thank you Jeff,
> Seems like time plays crucial role in HTM. This leads me to the question if
> it is implemented somehow differently or specially apart from HTM? Or it just
> simply works due to HTM nature? I am asking because HTM is memory and memory
> (something like RAM in your PC) has notion of HOW data are stored and not
> WHEN data was stored. Am I wrong?
>
> Thank you
>
>
>
>> On 01/11/2016 03:29 PM, Jeff Fohl wrote:
>> Wakan -
>>
>> I think you are getting there. In regard to your questions:
>>
>> 1. Spatial concepts are building blocks for temporal ones? - Yes,
>> essentially. One way of describing it is that "spatial" refers to
>> "things" and "temporal" refers to "things over time". So, your
>> "thing" could be a musical note, a word, or even an abstract
>> concept. A temporal series is simply these things put into a series.
>> 2. Yes, you could have data that is non-temporal, but it wouldn't be of
>> much use to NuPIC. You could also have data that is just time (say,
>> a bunch of timestamps), but that wouldn't be of much use either.
>> 3. Yes, encoders (within the context of HTM) will produce a spatial
>> representation of something, which then can be fed into an HTM model.
>>
>> I sense that you are getting hung up on the word "spatial". This is not
>> a NuPIC jargon word. It is more of a mathematics jargon word. Think
>> about how you might describe a physical object in "space",
>> mathematically. For a cube, you might describe its volume as X^3, where
>> X is the length of a side. Nowhere are you mentioning time, or how the
>> cube might change over time. In this situation, all you are concerned
>> with is with describing its static, non-temporal characteristics. Thus,
>> you are describing its "spatial" qualities. If you were to start
>> describing the cube's position over time as it moves through space (say,
>> after you give it a push), then you would be talking about its temporal
>> characteristics.
>>
>> I hope that helps.
>>
>> - Jeff
>>
>> On Sun, Jan 10, 2016 at 1:18 PM Wakan Tanka <[email protected]
>> <mailto:[email protected]>> wrote:
>>
>> Thank you Jeff,
>> this is clear example. May I ask regarding Matt's tutorials on youtube:
>> 1. sine waves
>> 2. hotgym prediction
>> 3. audio stream analysis
>> 4. geospatial tracking
>> 5. traffic anomalies
>>
>> what is spatial/temporal in those cases? I guess:
>> 1.
>> - temporal - the sines periods
>> - spatial - building blocks of those period
>>
>> 2.
>> - temporal - patterns representing day, week, year (those that you can
>> see repeating when you look at plot).
>> - spatial - building blocks of temporal
>>
>> 3.
>> - temporal - sequence of spatial
>> - spatial - depends on encoder (you might or not consider of changing
>> e.g. volume or instruments in audio I do not know what all was Matt
>> considering)
>>
>> 4.
>> - spatial - vectors
>> - temporal - everything that is composed of vectors
>>
>> 5.
>> - spatial - ???
>> - temporal - ???
>>
>>
>> Is this assumptions correct:
>> 1. From your post I've understood that spatial are building blocks for
>> temporal?
>>
>> 2. I guess that it is possible to have data where you have just spatial
>> data but not temporal (when you play chords at random), but not vice
>> versa?
>>
>> 3. Spatial is heavily depended on encoder and temporal are just logical
>> consequence?
>>
>>
>> Thank you
>>
>>
>>
>> On 01/10/2016 09:26 PM, Jeff Fohl wrote:
>> > Wakan -
>> >
>> > Perhaps a music analogy would be useful.
>> >
>> > Say you are playing a series of chords on a piano.
>> >
>> > Some of the chords are similar, some are not. Think of each chord
>> as a
>> > pattern. This is a "spatial" pattern. The term "spatial" perhaps is
>> > confusing because one tends to think of physical space when
>> hearing that
>> > term.
>> >
>> > "Temporal" refers to patterns over time. So, the sequence of the
>> chords
>> > is a "temporal" pattern.
>> >
>> > Does that help?
>> >
>> > - Jeff
>> >
>> > On Sun, Jan 10, 2016 at 12:16 PM Wakan Tanka <[email protected]
>> <mailto:[email protected]>
>> > <mailto:[email protected] <mailto:[email protected]>>> wrote:
>> >
>> > Thanks David,
>> > To be honest it did not help much :-( some example would be fine.
>> > Thank you
>> >
>> >
>> > On January 10, 2016 8:28:41 PM CET, David Ray
>> > <[email protected]
>> <mailto:[email protected]>
>> <mailto:[email protected]
>> <mailto:[email protected]>>> wrote:
>> >
>> > Hi Wakan,
>> >
>> > This definition may be useful:
>> >
>> > Spatial = the relative "nearness" of two data points in
>> terms of
>> > their semantic (characteristics of "meaning") attributes.
>> >
>> > Temporal (more accurately; Sequential) = refers to
>> patterns in
>> > "encounter order" pertaining to discrete units of input.
>> >
>> > Was that helpful?
>> >
>> > Cheers,
>> > David
>> >
>> > Sent from my iPhone
>> >
>> > On Jan 10, 2016, at 6:36 AM, Wakan Tanka
>> <[email protected] <mailto:[email protected]>
>> > <mailto:[email protected] <mailto:[email protected]>>>
>> wrote:
>> >
>> >> Hello Matt,
>> >> Can you please clarify differences between temporal and
>> >> spatial data? I've never really get into. Thank you very
>> much.
>> >>
>> >> On January 7, 2016 6:32:37 PM CET, Matthew Taylor
>> >> <[email protected] <mailto:[email protected]>
>> <mailto:[email protected] <mailto:[email protected]>>> wrote:
>> >>
>> >> Divyang,
>> >>
>> >> NuPIC can run on OS X and Linux (we test on Ubuntu), and
>> >> Windows (if you have your compilers installed properly,
>> >> see the READMEs).
>> >>
>> >> If by "local and global" anomalies, you mean
>> discrepancies
>> >> in short-term patterns vs long-term patterns, then yes.
>> >> For example, a change in hourly patterns would be
>> just as
>> >> anomalous as a change in daily patterns.
>> >>
>> >> Real-time or batch is supported, but the data should be
>> >> temporal in nature, not entirely spatial.
>> >>
>> >> ---------
>> >> Matt Taylor
>> >> OS Community Flag-Bearer
>> >> Numenta
>> >>
>> >> On Thu, Jan 7, 2016 at 12:14 AM, Divyang Shah via nupic
>> >> <[email protected]
>> <mailto:[email protected]> <mailto:[email protected]
>> <mailto:[email protected]>>>
>> >> wrote:
>> >>
>> >>
>> >>
>> >> ---------- Forwarded message ----------
>> >> From: Divyang Shah <[email protected]
>> <mailto:[email protected]>
>> >> <mailto:[email protected]
>> <mailto:[email protected]>>>
>> >> To: "[email protected]
>> <mailto:[email protected]>
>> >> <mailto:[email protected]
>> <mailto:[email protected]>>"
>> >> <[email protected]
>> <mailto:[email protected]> <mailto:[email protected]
>> <mailto:[email protected]>>>
>> >> Cc:
>> >> Date: Thu, 7 Jan 2016 08:08:15 +0000 (UTC)
>> >> Subject: some information for nupic anomaly
>> detection
>> >> hi,
>> >>
>> >> what are the different os and other configuration
>> >> support require for this project of anomaly
>> detection?
>> >> do you detect both local & global anomaly?
>> >> it supports only for real-time data or also support
>> >> anomaly detection for batch data?
>> >>
>> >> Thanks,
>> >> Divyang Shah
>> >>
>> >>
>> >>
>> >> --
>> >> Sent from my Android device with K-9 Mail. Please excuse my
>> >> brevity.
>> >
>> >
>> > --
>> > Sent from my Android device with K-9 Mail. Please excuse my
>> brevity.
>> >
>
>
