Matt Ginsberg's GIB bridge program lost to Zia and Rosenberg
in tournament play by 6.3 IMPs over 14 boards. (International
master points?) Ginsberg says that "at least on this day,
GIB deserved to be playing at the same table" with two of
the world's best bridge players. For equal players,
"the expected difference after 14 deals would be 20.6 IMPs,
simply because bridge is a stochastic game." (The expected
absolute difference between A and B is different from the expected
difference of A over B or of B over A, which would be zero.)
GIB has also been playing in the world computer championships,
where it defeated Bridge Baron by 83 IMPs in 48 deals
(with other rounds yet to play). In earlier play GIB was 6-0,
ahead of Q-Plus (4-2), Bridge Baron (4-2), Micro Bridge (3-3),
and three other programs. [,
comp.ai and rec.games.bridge, 28Jul98.]
SexNet is a neural network trained to guess the gender
of passport-style photographs (excluding hair, jewelry,
and clothing). A NIPS3 article by Golomb, Lawrence,
and Sejnowski says that SexNet achieved an accuracy of 91.9%
vs. human performance of 88.4%. [Will Dwinnell
<76743.1740@compuserve.com>, comp.ai.games, 30Jul98.]
(I don't know the details, but I'm suspicious. If the training
and tests sets had a sex-correlated difference of any sort,
it could easily explain such results. Suppose two different
camera/lighting stations were involved, for instance;
a neural network could easily spot a slight blur or 1% difference
in illumination, but a human could not. Women's makeup
could have a similar effect, as could ethnic differences
or any consistent difference in cropping of the hair, jewelry,
and clothes. Even if the system is accurate, it might be
easily confused by variable lighting or imaging conditions,
or fooled by anyone wearing unusual makeup. Still, this
sounds like a promising start.)
The Los Angeles Association for Computing Machinery
Technical Activity Committee on Artificial Intelligence (TACART)
is shutting down in Jun98, after years of seminar meetings.
There just aren't enough volunteers to keep it going.
[L. Andrew Campbell , comp.ai, 29Jul98.]
