[SL4] ESSAY: Data throughput
From: Barton, James (James.Barton@sweetandmaxwell.co.uk)
Date: Thu Jul 27 2000 - 03:49:09 MDT
Caution to the reader:
1. In all likelihood what I have written is nonsense.
2. I have done no research.
3. I don't even have a point to make. It's an indulgent ramble / brain-dump.
That said, I always wondered about attempts at AI I read of in the news.
Results were obviously never quite as wizz-bang as was hoped. Thinking about
human development, even up to a very young age, there was one apparent
difference between infants and experimental AIs (I recall no names or
details - see 2 above), that of throughput.
I've seen estimates of the bit rate through the optical nerves (can't
remember where, or what the figures were), and while I guess the other
sensory inputs don't have the same data rates, they must add up to a few
kilobits a second between them. Does anyone have a figure for any of those?
Up until the age of four, a child may be awake for, say, sixty million
seconds. At (guesstimate) ten kbps, that adds up to over 70 GB of data.
Obviously it isn't all retained, but some of it is retained, and some of it
is used to train various systems (motor, language, social, visual &c)
What I'm wondering is, how much ability to understand do you want to give an
AI to start with, and to what extent should it bootstrap its own
understanding? How well have other AI systems done with similar amounts of
data and more or less prewritten perceptual systems?
To hypothesise yet further, sight being our most important sense, our early
grasp of ontology is informed largely by it. A decent AI would presumably
know plenty about the world outside of itself, so the notion of
correspondence between internal reality models and external realities will
have to be suggested at some point (I'm still sceptical about an AI's
ability to select low-level goals itself - see my JOIN post for why).
The happenstance of evolution affords us a learning mechanism in the form of
a brain that can handle (and is driven to understand) certain specialised
work much more easily than other sorts. What can the study of infant
learning teach us about what we are born knowing, and can we see how much of
that would be directly useful? There's nothing logically necessary about
what evolution has favoured, but at least we already know it works.
I shall bring down the guillotine on this ramble now. Please - somebody
disagree with the contents of my JOIN post :)
James
This archive was generated by hypermail 2.1.5
: Wed Jul 17 2013 - 04:00:35 MDT