[info] [agi] Fwd: Numenta Newsletter: March 20, 2007

[Eugen Leitl]

----- Forwarded message from Eugen Leitl <eugen at leitl.org> -----

From: Eugen Leitl <eugen at leitl.org>
Date: Wed, 21 Mar 2007 13:50:34 +0100
To: agi at v2.listbox.com
Subject: Re: [agi] Fwd: Numenta Newsletter: March 20, 2007
User-Agent: Mutt/1.5.9i
Reply-To: agi at v2.listbox.com

On Wed, Mar 21, 2007 at 08:21:57AM -0400, Ben Goertzel wrote:

> Eventually, yeah, a useful AGI should be able to process visual info,
> just like it should be able to understand human language.

Being able to learn to see and to learn to hear, yes? How much
of it do you expect to be hardwired?

E.g. part of what cochlea does directly in hardware is a Fourier
transform. Do you expect to start with that as a prepositioned
building block, or let the system figure out the appropriate
transformation on its own?

> But I feel that the strong focus on vision that characterizes much
> AI work today (especially AI work with a neuroscience foundation)
> generally tends to lead in the wrong direction, because vision
> processing in humans is carried out largely by fairly specialized
> structures and processes (albeit in combination with more general-

There's a reason for that, it's in the law of optics and the
kind of structures a camera sees in the world. Vision (or an equivalent
high-bandwidth channel, direct depth perception by TOF, or LIDAR,
whatever) is a basic instrument for knowledge extraction.

You can skip on that by making the agent directly sense the simulation
grid (externalising the representation), but you'd have to abandon that
if your system does its first steps in the real world.

> purpose structures and processes).  So, one can easily progress 
> incrementally
> toward better and better vision processing systems, via better and
> better emulating the specialized component of human vision processing,
> without touching the general-understanding-based component...

Only parts of the visual processing pathway are "hardwired" (not really,
and it's not a linear thing at all), and of course the upper stages use
every trick the neocortex can muster. So, no, I don't think you can
trivialize vision, or postpone it.
 
> Of course, the same dynamic happens across all areas of AI
> (creating specialized rather than general methods being a better
> way to get impressive, demonstrable incremental progress), but it happens
> particularly acutely with vision
> 
> Gary Lynch, in the late 80's, made some strong arguments as to why
> olfaction might in some ways be a better avenue to cognition than vision.
> Walter Freeman's work on the neuroscience of olfaction is inspired by
> this same idea.

The bit rate you get from olfaction is really low. Yes, you can sense
gradients, and if you code everything with volatile carriers you can
recognize about everything. What I don't like about olfaction is that
it's evolutionary even older than vision, and directly wired to attention
allocation (emotion) processes. It's like vision, only without the
advantages, and even more hardwired.
 
> One point is that vision processing has an atypically hierarchical 
> structure in the
> human brain.  Olfaction OTOH seems to work more based on attractors
> and nonlinear dynamics (cf Freeman's work), sorta like a fancier Hopfield
> net (w/asymmetric weights thus leading to non fixed point attractors).  The
> focus on vision has led many researchers to overly focus on the hierarchal
> aspect rather than the attractor aspect, whereas both aspects obviously 
> play
> a bit role in cognition.

Makes sense.
 
> Direct "sensory" connections to biomedical lab equipment would
> be more useful ;-)

It would be interesting to see which sensory modalities are optimal
e.g. for medical voxelsets. One could wire a 3d retina on the voxelset,
or let the system look at the space/time domains directly, and let
it build its own processing and representation.
 
-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a> http://leitl.org
______________________________________________________________
ICBM: 48.07100, 11.36820            http://www.ativel.com
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE

-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

----- End forwarded message -----
-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a> http://leitl.org
______________________________________________________________
ICBM: 48.07100, 11.36820            http://www.ativel.com
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE