2010Q2 Discussions

[Phoenix]

I am so sorry I was away for a long time. Its taking me time to settle
down in USA. I am really not updated with the recent developments in
the project. I will catch up with you after reading all the wiki
stuff.

I am creating this new post for Q2 2010 discussions. I think we should
have a project plan specially when we missed all the Q1 deadlines. I
will try my best to catch up ASAP!!.

Regards,
Sarbjit

[Phoenix]

Got build errors while compiling the build.

Is it possible that we have a build server setup if there is any open
source available and you have some bandwidth on your system.

-sarbjit

[Habbit]

ok, np actually, I'm proceeding with my own pace.
Project as such has no deadlines.
Intermediate deadlines can just help to coordinate efforts.

If you start driving some area, planning will do have the meaning.

Let me explain what is happening now:
Q2 is away currently, Q3 is in the middle.
Q2 results are:

Migration to free IDE:
After trying several times to work with Codelight, especially with
debug - I decided to revert to Visual C,
I have downloaded VC 2010 Express Edition, which is free, registered
it and I'm now using it
I created solution Engine+Generic (still the same principles - copy
binaries).
All sources were committed.
Project files - are not, as still not clear what to commit and how to
work with it in version control.
If include all sources from all folders - it will be compiled
successfully.
Repeat folders in file tree in VC project.
Run is available now.

Auto-build and deployment - thrown away as not required

Libraries:
Changed my mind re principal approach - w/o solving practical things
it is impossible to prepare required libraries.
So I'm now trying to create AI engine as is and will add required
libraries when needed.
Not any library, but 2 types of libraries - Artificial Neural Networks
and Belief Networks.
Also I created multi-implementation framework - each type of library
can be represented by several implementations - from any source.
Also I'm confident that to move ahead with visible speed, the project
should re-use existing open source libraries w/o trying to rewrite
them as discussed before.
To reduce mess, current view is that facade interfaces should be
created for any inter-module call.

Test classes were detached from modules and combined within one test
module.
Test classes can reference any intra-module class directly w/o using
module interfaces, using private module include files.

Architecture:
Not sure it is final - but it is more plausible and practical compared
to previous one
Read aiengine\src\_wallpaper.txt
Project structure is almost the same as in architecture picture.

I'm now creating initial flow, starting from some primitive sensors.
I've implemented some starting model brain (not mind!) which has
ongoing focus in form of fixed set of networks, currently processed by
thread pool (all is configurable).
Inter-area connections, intra-area connections are my current interest
- with Hebbian learning.
Area is constructed and then processed (calculated, evolved) not by
brain component, but by dedicated mind component.

Again, the next question is how to convert sensing data into
(internal) features.
So HTM is still actual thing - but only in the context of real tasks.
Let' name it continuous integration.

Please let me know your thoughts and area.

[Phoenix]

If we reverted back to Visual Studio than we need to delete workspace
files related to codelite from src. I think its good decision as we
might end up spending more time on struggling with the IDE problems
and limitations.

I will make sure that is I am writing any library than I will provide
a interface to the library and hide all the complexities from other
modules.

We need to make sure that libraries are only accesses using interfaces
and no direct call to any function is made from any module.

I will create a development environment using visual studio on my
machine and will let you know.

[Phoenix]

Can you please mail me or check in the project files too as I want to
be consistent with the project configuration with you.

[Habbit]

Commited:

C:\projects\ai2svn\human\workspace.human>svn commit -m "add VC 10
files"
Adding workspace.human\aiengine.vc
Adding workspace.human\aiengine.vc\aiengine.vc.vcxproj
Adding workspace.human\aiengine.vc\aiengine.vc.vcxproj.filters
Adding workspace.human\generic.vc
Adding workspace.human\generic.vc\generic.vc.vcxproj
Adding workspace.human\generic.vc\generic.vc.vcxproj.filters
Adding workspace.human\human.solution.vc
Adding workspace.human\human.solution.vc\human.vc.sln
Transmitting file data .....
Committed revision 699.

[Phoenix]

Thank you for the project files. Solution seems to be building fine.
Just faced one problem API 'rfc_sys_getcpuload' in modbrain-
>activememorythread.cpp was not implemented anywhere. Kindly look into
it.

As you were actively driving this project please guide me about module
libbn. I believe you want me to work in this module. I saw some code
already written for that. Can you please let me know where exactly is
this module now.

Kindly also help to clarify the path that I should follow to proceed
with that module.

Thanks.

[Phoenix]

Also I think we should have a regular meeting everyday for may be 30
minutes where we can clarify and discuss something about the project.
I can make myself available anytime according to your availability.
Let me know your views about it. It will help align our thoughts to
the right direction.

Thanks

[Habbit]

1. re link error:

sorry, forgot to commit last generic files - just done it:

C:\projects\ai2svn\human\generic\src>svn commit -m "new functions"
Sending src\include\_genos.h
Sending src\modos\_gen_win.c
Transmitting file data ..
Committed revision 700.

2. re libbn

If you try to focus on HTM-related things, I would recommend to pay
attention to modcognition/neocortex component as per
http://code.google.com/p/ahuman/wiki/Architecture
In architecture, this component can use one or more libraries from
libbn, which, in turn, is a set of Belief Network implementation
libraries.
Architecture allows multiple BN implementations - so you are free to
start any number of libraries there.
Existing libbn\view_custom library was trial attempt to process visual
data in realtime.
It is up to you to kill it or to rework to something useful.

The challenge is to create working part of mind process - perception
(part of cognition), not just some helper library.

HTM article I have used for implementation answers the question how to
reduce big image data to one value which is related to the "name" of
object represented by image.
It is not the case for perception - where sensor data. e.g. visual
image triggers neurons corresponding to various "features" and their
abstractions (for instance, angle/thin angle/thick angle).
If you look at real human mind, you'll find it looks like pyramide
with top in visual cortex and bottom deeper. So number of neurons
triggered by sensor data is greater than number of sensor neurons.

Also, we have "sensor signal" (streaming sensor data - snapshots of
state of sensor neurons), "feature", "generalised feature" (do we have
this really?), "object", "generalised object".
Current understanding is that associations give us object items, while
neocortex - only feature items.
Al these questions and answers better be collected on releated
research page with some plausible functional model - at
http://code.google.com/p/ahuman/wiki/MemoryPredictionResearch.

Please create kind of working neocortex which:

- has inputs from inter-area (Sensors->NeoCortex) connection streamed
by IO channel "sensordata"
- this channel will contain events referencing neural networks inside
neocortex triggered by sensor data updates

If you have some good open source library - I'd suggest to commit it
to libbn/xxx and start using it.

Still open questions:

- how these connections to specific networks are established and how
they evolve
- what exactly is output of neocortex (now it is named as "features")

Until I finish my sensor data work, you can have mock implementation
where you generate channel data on "sensordata" channel from some new
thread or with test requests w/o actual involvement of sensors.
Then we can integrate our results.

I started sub-page of http://code.google.com/p/ahuman/wiki/MemoryPredictionResearch
- http://code.google.com/p/ahuman/wiki/InternalRepresentationResearch.
Please start draft content there - I will add details on page when
required.

[Phoenix]

Ok I now have a sight as to which area should I focus on. Thank you
for clarification and I am going through the code to see the
architecture implementation.

I am also doing some research on the perception models and updating
the related research pages in wiki. If you have some time I want you
to take a look into this link.

http://ilab.usc.edu/toolkit/documentation.shtml

It seems a library that we can customize according to out needs. It
implements the salients maps for visual recognition and sends these
signals to brain same as that happens in biological brain.

Let me know your thoughts about this.

Regards,
Sarbjit

On Aug 17, 3:29 pm, Habbit <vsavc...@gmail.com> wrote:
> 1. re link error:
>
> sorry, forgot to commit last generic files - just done it:
>
> C:\projects\ai2svn\human\generic\src>svn commit -m "new functions"
> Sending        src\include\_genos.h
> Sending        src\modos\_gen_win.c
> Transmitting file data ..
> Committed revision 700.
>
> 2. re libbn
>
> If you try to focus on HTM-related things, I would recommend to pay

> attention to modcognition/neocortex component as perhttp://code.google.com/p/ahuman/wiki/Architecture

> research page with some plausible functional model - athttp://code.google.com/p/ahuman/wiki/MemoryPredictionResearch.

>
> Please create kind of working neocortex which:
>
> - has inputs from inter-area (Sensors->NeoCortex) connection streamed
> by IO channel "sensordata"
> - this channel will contain events referencing neural networks inside
> neocortex triggered by sensor data updates
>
> If you have some good open source library - I'd suggest to commit it
> to libbn/xxx and start using it.
>
> Still open questions:
>
> - how these connections to specific networks are established and how
> they evolve
> - what exactly is output of neocortex (now it is named as "features")
>
> Until I finish my sensor data work, you can have mock implementation
> where you generate channel data on "sensordata" channel from some new
> thread or with test requests w/o actual involvement of sensors.
> Then we can integrate our results.
>
> I started sub-page ofhttp://code.google.com/p/ahuman/wiki/MemoryPredictionResearch

> -http://code.google.com/p/ahuman/wiki/InternalRepresentationResearch.

[Habbit]

After reading iLab presentation my thoughts are:

- it is mature library, grounded quite before HTM - can be powerful
but stale in terms of techonology
- it is video signal processing system, with "object" outputs
- it is toolkit with high resource requirements
- possible targets we have seen before for aHuman will not use vision
sensor - see http://code.google.com/p/ahuman/
- from above - it could be trial research stuff, but not for final use

It is up to you, my view is that better to find more universal and
recent library.

> ...
>
> продолжение »

[Habbit]

After looking into iLab presentation:

- it is about visual processing - probably our possible targets will
not have such sensors - see http://code.google.com/p/ahuman/
- it is mature but old enough, pre-HTM
- looks like many resources are required

So it's up to you, but I think better to find more universal and
recent library for final implementation.
For research purposes it can be used though.

[Habbit]

just noticed - "Also I think we should have a regular meeting everyday

for may be 30
minutes where we can clarify and discuss something about the project.
I can make myself available anytime according to your availability."

Agree, it can help us to understand each other. I'd suggest start some
exchange the day after day, e.g. each even day of month around 8 PM
GMT.
Let me know your time zone and we can find suitable time.
I have spare time around 7PM-10PM GMT, a bit more flexible on
weekends.

[Phoenix]

I am on EST (Eastern Standard Time) and I am free form 5 pm to 6 pm, i
think it will be 9PM - 10 PM GMT, correct me if I am wrong. I think
that time will be best. Let me know your views.


-Sarbjit

[Habbit]

your time = EST = GMT - 5
my time = GMT - 4 in Summer = your time + 9

It means you are available 2-3AM in my time.
I go to bed around 1-2 PM.
Hence it is not quite comfortable time window.

Maybe it is just calculation mistake.
Let me know time when you have received this message.
It was sent at 10PM my time.

[Phoenix]

It was 2 PM my time and I think its quite ok we can have discussion
from 1pm - 2pm my time( 9 pm -10pm) your time. I will manage it. Let
me know your views

Thanks