New FKI-Reports

[Juergen Schmidhuber]

Two new reports on spatio-temporal credit assignment in neural networks 
for adaptive control are available.

   LEARNING TO GENERATE FOCUS TRAJECTORIES FOR ATTENTIVE VISION
	                FKI-REPORT 128-90
               Juergen Schmidhuber and Rudolf Huber

One motivation of this paper is to replace the often unsuccessful 
and inefficient purely static `neural' approaches to visual pattern 
recognition by a more efficient sequential approach. The latter is 
inspired by the observation that biological systems employ sequential 
eye-movements for pattern recognition. 
The other motivation is to demonstrate that there is at least one 
principle which can lead to the LEARNING of dynamic selective 
spatial attention.
A system consisting of an adaptive  `model network' interacting with 
a dynamic adaptive `control network' is described. The system LEARNS
to generate focus trajectories such that the final position of a 
moving focus corresponds to a target to be detected in a visual scene.  
The difficulty is that no teacher provides the desired activations of 
`eye-muscles' at various times. The only goal information is the 
desired final input corresponding to the target. Thus the task 
involves a complex temporal credit assignment problem, as well as an 
attention shifting problem. 
It is demonstrated experimentally that the system is able to learn
correct sequences of focus movements involving translations and
rotations. The system also learns to track a moving target. 
Some implications for attentive systems in general are discussed. 
For instance, one can build a `mental focus' which operates
on the set of internal representations of a neural system. It is 
suggested that self-referential systems which model the 
consequences of their own `mental focus shifts' open the door for 
introspective learning in neural networks. 


        TOWARDS COMPOSITIONAL LEARNING IN NEURAL NETWORKS 
	                FKI-REPORT 129-90
                       Juergen Schmidhuber 

None of the existing learning algorithms for neural networks with 
internal and/or external feedback addresses the problem of learning 
by composing subprograms, of learning `to divide and conquer'. In 
this work it is argued that algorithms based on pure gradient descent 
or on temporal difference methods are not suitable for large scale 
dynamic control problems, and that there is a need for algorithms 
that perform `compositional learning'.  Some problems associated with 
compositional learning are identified, and a system is described which 
attacks at least one of them.  The system learns to generate sub-goals 
that help to achieve its main goals. This is done with the help of 
`time-bridging' adaptive models that predict the effects of the 
system's sub-programs. A simple experiment is reported which demonstrates 
the feasibility of the method.


To obtain copies of these reports, write to 

Juergen Schmidhuber
Institut fuer Informatik,
Technische Universitaet Muenchen
Arcisstr. 21
8000 Muenchen 2
GERMANY

or send email to
schmidhu at lan.informatik.tu-muenchen.dbp.de

Only if this does not work for some reason, try
schmidhu at tumult.informatik.tu-muenchen.de

Please let your message look like this:



subject: FKI-Reports
FKI-128-90, FKI-129-90
Physical address (not more than 33 characters per line)




DO NOT USE REPLY!