Let's return, for a moment, to our hidden
agenda, of trying to understand something
about intelligence from all the stuff that
we've learned so far.
We figured out that classes can be learned
from experience.
Features can also be learned from
experience.
For example genres,
That is classes as, as well as roles which
can maybe a features merely from the
experience of people buying books.
What is the minimum capability needed to
learn features and classes directly from
data?
This is a, rather carefully thought out
question,
So let me think about it for a minute.
The first stage one needs some low level
of perception,
One needs to be able to perceive in the
case of humans, pixels and frequencies or
in the case of our systems, one needs to
know, be able to identify the person by a
person ID, the book by a book ID and
that's about it.
Second, one needs the ability to subitize
which is another way of saying counting or
distinguishing between one and two things.
So it turns out that very young babies are
actually able to distinguish between one
person or two people,
One object or two objects and they get
surprised when suddenly one object
disappears from the scene,
So this is essentially something innate.
Similarly, the ability to break up
temporal experiences that one experiences
over time into episodes that they
experienced something in the past five
minutes and then the next ten minutes
another experience because suddenly,the
scene has changed.
To break up this episodes is another
subitizing feature in time, which babies
learn at a slightly later age.
Given these two things, and our hidden
latent model techniques, one can
essentially in principle learn classes and
features together simply from the fact
that they co-occur together in
experiences.
Theoretically it works,
But in practice, lots of research is
currently underway to enable machines to
learn in an unsupervised manner, both the
classes as well as the features.
So you're clustering the classes, you're
clustering the features side-by-side,
using the fact that classes and features
co-occur in different experiences or
objects to learn both together.
So these is really at the frontier of
research today in both from web
intelligence as well as understanding
human intelligence, to a certain extent.
So when you come across articles which
talk about bottom up learning or grounded
techniques, essentially they're talking
about things like this where one is trying
to learn a hidden or latent model directly
without supervision.
Of course one might really ask, to what
extent have we actually learned anything
in the true or pure sense of the word.
In fact, in a rather celebrated 1980
article the Philosopher, Philosopher John
Searle refuted any suggestion that
mechanical reasoning using all manner of
learned facts or, Or, or rules could
actually be considered intelligent and
this is the argument he used reminiscent
of the Turing test, in fact.
Sir imagined a room where a person,
In this case, it's a bird,
But he imagined a person armed with rules
and facts and reasoning techniques that
allowed that person to translate from
Chinese to English using mechanical
calculations and facts and rules about how
to translate.
The question, so the last was does the
translator know Chinese in the sense that
a native speaker of Chinese knows Chinese.
He argued vehemently that this person
could not in any sense be construed to
know Chinese, and therefore the prospect
of machine intelligence divorced from any
direct perception or abilities in,
Say the language, Chinese was actually a
fallacy.
Interestingly, exactly such techniques
such as we have discussed here as well as
new ones that we'll talk about next time
like hidden Markov models, are actually
used to parse and translate Chinese into
English fairly well today in Google
Translate.
The question posed by Sir has now become
popularly and known as the Chinese room
debate and is certainly worth pondering
about, when one talks about web
intelligence as having taught us anything
about our own abilities.