[FOM] Fuzziness

Sandro Skansi skansi.sandro at gmail.com
Wed Jul 20 03:28:17 EDT 2016


Dear professor Friedman,

About you question:
"Also, how does it relate to recent breakthroughs in machine learning,
deep leaning, etcetera?"

I will speak from primarily an engineering point. Deep Learning is deeply
connected with fuzzy logic, but they are also distinct. The current
approach is to use DL for classification, and due to computational
cheapness, delegate as much as one can to this classification, and then use
fuzzy systems to reason with learned predicated. In fact, one could view it
as:

state-of-the-art AI system = Deep Learning + Fuzzy Logic

Say you run a book store, and you want to send a complimentary birthday
book to each client, but you only have their ID and purchase history. What
you need is a classifier that puts the client in categories. Say, "young
female linguist", and sends the "young-female-linguist" book. But, as
classification comes with a confidence score, of say 0.25 (to be
interpreted as percentage), and you are not sure. You want the system to
act only on high confidence (fuzzy set membership), and to look into the
components (and choose just the components with a confidence above a
certain level) if the confidence is below a certain level. This kind of
systems can get quite complicated (in terms of fuzzy inference) quite fast,
and it seems easier to try to "push" the complicated parts in DL, since
they are capable of improvement with new data.

I think the next big thing would be to emulate a system similar to
professor Pedro Domingos Markov Logic Networks, but for a system of fuzzy
logic, so that the inference system itself, not just the DL parts, can
change and adapt.

Cheers,
Sandro


### check out my profile and publications at
https://hr.linkedin.com/in/skansi ###


On 22 June 2016 at 03:07, Harvey Friedman <hmflogic at gmail.com> wrote:

> On Tue, Jun 21, 2016 at 4:21 PM, Kreinovich, Vladik <vladik at utep.edu>
> wrote:
> > There is a whole direction led by Dr. Scott Dick, google "complex-valued
> > fuzzy logic" and you will find many links to related papers, there are
> > regular sessions at fuzzy conferences on this topic. You may want to ask
> > such questions to fuzzy mailing lists instead of Foundations of Math one
> >
>
> Except that the very idea of doing "fuzzy logic" is deeply
> foundational, and it could be of great interest for the FOM to see a
> modern account of the basics of fuzzy logic and fuzzy mathematics and
> fuzziness from the ground up. it would seem likely that f.o.m. people
> who have not been informed much about it would have something
> interesting and/or useful to say about it.
>
> Also, how does it relate to recent breakthroughs in machine learning,
> deep leaning, etcetera?
>
> Harvey Friedman
> _______________________________________________
> FOM mailing list
> FOM at cs.nyu.edu
> http://www.cs.nyu.edu/mailman/listinfo/fom
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: </pipermail/fom/attachments/20160720/8677d8dc/attachment.html>


More information about the FOM mailing list