Oesterreichisches Forschungsinstitut fuer Artificial Intelligence (OeFAI)
Schottengasse 3, A-1010 Wien
Tel.+43-1-53532810,5336112
-------------------------------------------------------------------------
VORTRAG
*******
Dr.Peter TINO
Slovak University of Technology
Bratislava
Neural Modeling of Complex Symbolic Sequences
---------------------------------------------
We study the problem of modeling long, complex symbolic sequences
with recurrent neural networks and stochastic machines. The talk
will be focused on the following issues:
1. What do we mean by saying that a long symbolic sequence is
complex?
2. How to monitor the training process in which recurrent networks
are trained on complex sequences? (We use information theory
based performance measures.)
3. How to extract the knowledge the recurrent net has gained during
the training process? (We use stochastic finite state machine
representations.)
4. How do we know the extracted machine represents the original
network well?
5. We shall present a tool for an "intelligent" transformation of
a sequence temporal symbolic structure into a spatial fractal
representation and present theorems supporting the claim that
this transformation is "intelligent".
6. We show how to build context dependent variable length "Markov
systems" from spatial representations of temporal sequence
structures (analogical to approaches of Ron, Singer and Tishby).
Zeit: Mittwoch, 16.Juli 1997, 18:30 Uhr pktl.
Ort: OeFAI, Schottengasse 3, 1010 Wien 1.