On Mon, Apr 8, 2013 at 8:27 PM, Ben Goertzel <[email protected]> wrote:
> On the purely vision side, it seems we can get a big enough DeSTIN hierarchy
> running on a recent-model GPU card (say, an Nvidia Kepler ...) ...
>
> On the symbolic side, I'm less sure, but I'm guessing that if the Atomspace is
> optimized (as Linas is planning) we could do it on a single
> contemporary server with,
> say, 6 or 8 processors and maxed-out RAM.....  There are no rigorous
> calculations
> for this, though, it will have to be determined via experiment

Yes, but how many nodes and links will that handle? How many do you
need for reading text or recognizing faces?

I earlier estimated for full human level vision, 10 layers of 10^6
nodes each, fully interconnected between layers with 50 ms response
time. That's 200 TFLOPS. I figure 10^6 inputs from the retina and 10^6
output visual concepts (10x vocabulary). I figure 50 ms because when I
look at a picture of several people, it takes me 0.5 seconds to
recognize each face.

Then you need about a decade's worth of video to train the system. You
probably don't want to wait a decade for each experiment, so you
really want something faster than 200 TFLOPS.

> I am unsure why you call SIFT a model of a fovea..
>
> We have discussed giving the DeSTIN hierarchy greater resolution in
> the center of the grid, which would be a clearer and simpler model of a fovea,
> I think...

I only saw a brief mention of SIFT. I thought that was the idea, but
also to move the fovea around to parts of the image with high
information content (or at least high spatial and temporal contrast).
Obviously this can be done in software, unlike the human eye. Either
way, I think it is a significant optimization (which I included in the
above estimates).

--
-- Matt Mahoney, [email protected]


-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to