Re: [Heartlogic-dev] RE: FW: Research in emotional AI

[Joshua N Pritikin]

On Thu, Mar 11, 2004 at 08:42:34AM -0800, Josh White wrote:
> > > If you still believe that any of these "smart algorithms" are really
> > > smart then I strongly encourage you to implement them in C.
> 
> I don't believe any existing algorithm are really smart, in the sense
> that they could even remotely compete with a human brain.  
> 
> The "you can only understand it if you build it" answer sounded
> defensive to me. 
> I do think building something makes sense if one has a deep or enduring
> an interest in the field.  Since I had already told you I'm a
> semi-layman I assumed you knew that I'm talking to experts like you
> because I'm looking to learn about the field generally, and get a broad,
> reasonably accurate understanding of the current state of AI.  In that
> context, it sounded like a defensive answer to me.

Well, if you put it like that then I have to retreat to Bill's
position --- we want to learn about psychology and we'll use any tool
available (including neural nets) which helps us reach our goal.

> Personally, I'm happy if I help (even slightly) create an advance in AI
> that could be applied in the next 2 years to greatly improve a product
> or service normal people want or need.  I think lots of scientific
> problems - say emulation of human emotion - is great, very applicable to
> lots of real-world problems.  I think AI is currently not there, but
> that it could be soon (but I'm not sure - again, hence the queries).  

Generally speaking, I totally agree.

> As you've seen, I'm interested in neural nets.  I understand neural nets
> on a shallow level. I've used a few simple neural net demos and
> understand how weighting a simple static neural network simulates basic
> neural pathway construction and thus crude "learning".
>       I don't understand the limits of neural nets. For example, it
> seems obvious to add self-modifying weighting to any number of other
> attributes of the data in the neural net, such as timing of signals,
> number of connections, types of signals transmitted, etc. I don't know
> why that doesn't work (if it's been done, which I assume it has).  Do
> you think that neural net could never be theoretically equivalent in
> functionality to a human brain?

My opinion is that it boils down to this: It helps a lot to use the
right tool for the job.  Given the problem of adding two 16 bit
integers, compare the solutions using digital logic gates and using
neural nets.  Then look at the solution in C: "i1 + i2".

That's why it's so important to develop an intuition for what kinds of
problems a particular algorithm is good at solving.  The same thing I
was trying to say in the previous few emails.  If you think that
neural nets are "interesting" then you can play around with them and
probe their limits.

>       Perhaps these questions are too basic to interest you. If so,
> can one of you point me to a good 'advanced beginner' guide to neural
> nets and other learning algorithms - ie a good deep FAQ?

That's a tough question, but let me give it a try ...

I think you can develop a reasonable intuition just by thinking about
the algorithm backwards --- What kinds of problems is this algorithm
designed to solve?  How robust is the algorithm against noise in the
training data?  What kinds of solutions does a given algorithm
produce?  Is there a tendency to under-fit or over-fit?

We need these questions because I don't think there is universal
algorithm (besides "go for it!").  I agree with Erik Mueller (author
of ThoughtTreasure), when he speculates that artificial intelligence
is like archeology.

So I think these kinds of questions help you build an intuition about
algorithms.  Then when you come across a particular kind of problem,
you can make a good guess at which algorithm is a most suitable
"reaction."  Hey, maybe we can build a neural net for that?  ;-)

-- 
A new cognitive theory of emotion, http://openheartlogic.org

signature.asc
Description: Digital signature