Anurag.Acharya@CENTRO.SOAR.CS.CMU.EDU (05/01/88)
THEORY/AI SEMINAR
Jiawei Hong, Courant Institute
Friday, May 6
2:00 p.m.
4605 WEH
Connectionist and Other Computational Models in AI
This talk consists of three problems:
1. The well known connectionist models defined by an nxn real weight matrix
(notice that an arbitrary real number may have infinite information in it),
can be simulated by a non-uniform circuit of O(n^3 log n) Boolean gates
(thus the total information is finite) with time slowdown O(log n). Therefore
the connectionist model do not have more computational power than other
parallel computation hardwares.
2. In the future, computers may consist of digital elements as well as analogue
elements. Which kind of analogue element does help? We prove that a kind of
analogue element does help only if
(1). the analogue element has very very high precision: exponentially many
significant bits, or
(2). the analogue element can very efficiently compute a problem which is
NOT in NC. Both are unlikely true.
3. Can human brains be simulated by computers in the future? Under some rea-
sonable assumptions, we proved that this is possible. (there is no intrinsic
difficulty, like NP-hardness.) We are optimistic.