2010Q1 discussion
Habbit
Project page containing current status is http://code.google.com/p/ahuman/wiki/ReleasePlanning.
Phoenix
through it. Kindly let me know what is the status as well as vision
for the Q1 2010.
Kindly elaborate as much as possible so that I can have better
understanding of what is going on in your mind.
Habbit
goal - only some trial of the code. Moreover, the concept behind HTM
(except core idea) is not clear.
I have checked my local storage of external theory/resedarch docs into
$/research/articles. It contains "2005 05 Analysis and Implementation
of MPF.pdf" doc - which is most specific about details.
Still I did not check sourceforge resources for the same.
Questions that exist now:
1. What is the output of HTM. Above article states it is 1 variable,
containing the "name" of recognized entity.
It has no much sense for us - as we do not create narrow-functional
device.
Habbit
In fact the brain when processing recognition outputs from sensor,
"reads" not the one top item of neocortex, but some top part of HTM
tree, to have reaction to set of recognized features.
2. Attention Focus mechanism is not implemented.
3. Cache is too slow for HTM - it is very important there.
The list could be extended, but I can say that currently I have very
basic understanding of HTM - it is very vague how it works and what
features it has - and its role for human project (also because AI
model is not created yet).
We need to have AI modules map with their primary function to
understand what is a place for HTM - see table at the end of
http://code.google.com/p/ahuman/wiki/Architecture.
Basic idea for 2010Q1 was described -
a) have all basic libraries working - modnnlib, modhtmlib, modbnlib?
(belief networks), modmplib? (markov process) (not rich, not best, but
fulfilling primary function)
b) have finalized architecture when list of modules, role and
technology behind all of the modules are clear and described (i.e.
transparent and reasonable model)
(also it all pages in the Wiki framework better to have their optimal
structure and level of details to perform coordination role)
What I'm doing now:
1. Coping with wxWidgets to finish Codelite migration - for Htmview
test app.
2. Reading various articles about emotions in AI - and thinking about
concepts for personality.
3. Trying to understand what is pragmatic target for 2010Q1.
Phoenix
Going through all the HTM related pdfs.
Will let you know my views of the same.
On Dec 28, 12:15 pm, Habbit <vsavc...@gmail.com> wrote:
> continued...
>
> In fact the brain when processing recognition outputs from sensor,
> "reads" not the one top item of neocortex, but some top part of HTM
> tree, to have reaction to set of recognized features.
>
> 2. Attention Focus mechanism is not implemented.
> 3. Cache is too slow for HTM - it is very important there.
>
> The list could be extended, but I can say that currently I have very
> basic understanding of HTM - it is very vague how it works and what
> features it has - and its role for human project (also because AI
> model is not created yet).
> We need to have AI modules map with their primary function to
> understand what is a place for HTM - see table at the end ofhttp://code.google.com/p/ahuman/wiki/Architecture.
Phoenix
going through the reading material and also if possible do create a
doc for sharing responsibilities in certain area so that we are clear
what we intend to do in what time frame...
Habbit
pooler are merged into one mixed pooler.
Second major bug is that recognition should occur only if prediction
has confirmed recognized sequence.
great resources as well - at http://sourceforge.net/projects/neocortex/
and http://mitworld.mit.edu/video/316 (generally you can google for
Video re Jeff Hawkins)
re doc: I think we can limit this to 2 pages:
- http://code.google.com/p/ahuman/wiki/ProjectPlanning - with some
"stairway to the heaven" which is seen more fine-grained in the
nearest phases and the sequence itself is not mapped to quarterly
timeframes - as it is seen currently with most detail (still high
level)
- http://code.google.com/p/ahuman/wiki/ReleasePlanning - which is
about current quarter and which is subject to freeze
So I propose we complete ReleasePlanning and have not contradicting
contents in ProjectPlanning - both agreed by the end of this week.
Phoenix
bugs once i am done with htm theory. Regarding the documents about
release planning and the project planning lets freeze the release
planning the freeze it.
PS: I am currently in USA and adjusting so will take some time to
settle..still i am working on the project so dont mind if there is no
response from my side...be assured that i am on..:)
On Jan 11, 2:14 pm, Habbit <vsavc...@gmail.com> wrote:
> btw, re Htm - one major design bug is that spatial pooler and temporal
> pooler are merged into one mixed pooler.
> Second major bug is that recognition should occur only if prediction
> has confirmed recognized sequence.
> great resources as well - athttp://sourceforge.net/projects/neocortex/
> andhttp://mitworld.mit.edu/video/316(generally you can google for
> Video re Jeff Hawkins)
>
> re doc: I think we can limit this to 2 pages:
>
> -http://code.google.com/p/ahuman/wiki/ProjectPlanning- with some
> "stairway to the heaven" which is seen more fine-grained in the
> nearest phases and the sequence itself is not mapped to quarterly
> timeframes - as it is seen currently with most detail (still high
> level)
> -http://code.google.com/p/ahuman/wiki/ReleasePlanning- which is
Habbit
you have time to do something else.
Are you alone there or have some support from relatives or friends?
Any case, good luck!
Re project - ok, let's limit to what is currently listed on planning
pages.
To split the ownership somehow, I would suggest you to finish with
htm, while I will look into automated Build/Deployment and
Reinforcement Learning.
Component design can be our shared task.
Reflected that in owner column.
Let me know if any suggestions.
One more thing - table has a column for implementation review mark.
I would like if we set this whenever we understand changes done are
clear and agreed.
I will review your changes, you - mine.
Whenever done - reviewer has to add his sign off there.
On Jan 17, 9:54 pm, Phoenix <sarbjit.lon...@gmail.com> wrote:
> I am currently looking at the htm documents and will look into the
> bugs once i am done with htm theory. Regarding the documents about
> release planning and the project planning lets freeze the release
> planning the freeze it.
>
> PS: I am currently in USA and adjusting so will take some time to
> settle..still i am working on the project so dont mind if there is no
> response from my side...be assured that i am on..:)
>
> On Jan 11, 2:14 pm, Habbit <vsavc...@gmail.com> wrote:
>
> > btw, re Htm - one major design bug is that spatial pooler and temporal
> > pooler are merged into one mixed pooler.
> > Second major bug is that recognition should occur only if prediction
> > has confirmed recognized sequence.
> > great resources as well - athttp://sourceforge.net/projects/neocortex/
> > andhttp://mitworld.mit.edu/video/316(generallyyou can google for
> > Video re Jeff Hawkins)
>
> > re doc: I think we can limit this to 2 pages:
>
> > -http://code.google.com/p/ahuman/wiki/ProjectPlanning-with some
> > "stairway to the heaven" which is seen more fine-grained in the
> > nearest phases and the sequence itself is not mapped to quarterly
> > timeframes - as it is seen currently with most detail (still high
> > level)
> > -http://code.google.com/p/ahuman/wiki/ReleasePlanning-which is
Phoenix
laws are here and my wife already had place here. I could not do much
in previous days as I was looking for a job...but i am always thinking
and studying abt the project as I am geared up for this project..:)
I have gone through the online material about htm and have fair idea
about the concept. I am still looking into more details as how we can
use it in out project effectively.
> > > andhttp://mitworld.mit.edu/video/316(generallyyoucan google for
Habbit
Hope you will find smth good.
How many open vacancies in US IT now?
I have just finished to read book Introducing Artificial Intelligence
by Henry Brighton.
It confirmed many my thoughts and has given some push for others.
I'm going to start filling AI Research page to reflect this
understanding.
Re Htm as it is probably one of fundamental technologies, maybe we
need to start research page as well.
It can be useful to store some facts and ideas as a permanent
information, not task-level.
> > > > andhttp://mitworld.mit.edu/video/316(generallyyoucangoogle for
Phoenix
concepts about HTM from Numenta's website. I am going to write some
code to test NuPic. Lets see what can I do. Numenta is free for
experimentation but not for commercial. Do you think it will be good
idea to use Numenta's NuPic.
Habbit
it can be a problem.
Also re 3d party AI libs I think we'd better to rewrite them line-by-
line - to be sure we understand how it works and be able to amend when
required.
For research purposes it is not a problem to use 3d party as is.
Also Numenta is relying more on belief network approach - which tends
to be more symbolic AI than neural-like (e.g. I don't like sending to
the top belief distribution).
Above are only thoughts, not limitations of any kind.
btw, I have created http://code.google.com/p/ahuman/wiki/MemoryPredictionResearch
- fill there whatever you find interesting.