What we learnt from EANN '97

Fredrik Norlund, ABB Industrial Systems AB, Sweden
and 
Abhay Bulsari, AB Nonlinear Solutions OY, Finland

Like  the previous two years, there  were many papers on various kinds
of applications of neural   networks in many engineering  fields. This
year, we also  had an industrial  panel discussion,  which was a  good
exchange of ideas and views regarding  the use of neural networks. The
industrial panel consisted of nine panel members from Sweden, Germany,
Brazil and USA. We intend to have a similar panel discussion next year
as well.  Out  of the  hundred-plus  participants, over  20% were from
industries. Most of  them were from  process industries and automation
companies.

There is a clear trend towards better training algorithms. At the same
time, people talk  less  and less about  training  algorithms in their
papers, which is a positive sign of progress.

It was pointed out that there is very little talk about the quality of
data. Most  papers mention  the amount  of data available.  However, a
large amount of  data is not worth very  much if it contains copies of
the same  observation with varying amounts of  noise. It  is therefore
important to take into  consideration the variation of important input
variables in the data, and the characteristics  of the noise contained
in the  data  set. The industries  which  are interested in developing
advanced  models   could think of    collecting  data appropriately to
facilitate the development of neural network models from it.

It was noticed  that there is  a trend towards hybrid  systems. Nobody
wants to throw away  physical models  developed  after years  of work,
even if they are  not as good as a  small network. Smaller networks in
combinations with physical models seem to be favoured.

There was a suggestion that some industries should put their data on a
www site for a competition. The participants of that competition could
be asked to perform a modelling  task using neural networks, and other
techniques. The idea is good and hopefully, some industry will be able
to do so.

You need the  right kind of  software  for solving your problems,  not
necessarily an expensive one.  There is some  tendency to believe that
buying the right software will solve  your problems. Software does not
solve real  problems; it   can  only train neural  networks  with some
algorithms. The  key   to  success lies  in  factors  other  than good
training of neural  networks.  Experience with solving   real problems
with neural    networks is useful,  and understanding   of the problem
domain also helps significantly.

There is   still some tendency  left to  feed  in everything to neural
networks  and expect  a two hundred   input network  to take  only the
necessary information from the input vector.

80% of the time  is spent in   peripheral issues, not in  training the
networks. This has been heard in many other places.

Neural networks are not blackboxes if you know how to analyse them. It
is true that they are not as  transparent as linear models or physical
models, but neural networks can also be analysed to some extent.

Most of the industries do not have specialised knowledge to be able to
decide the right kind of neural networks or alternative techniques for
a given problem.

Adaptation of the model to long term changes in processes was remarked
to be a problem with neural network modelling. It was then pointed out
that the problem is similar with other empirical models also.

The results should be available  to the operators  such that they feel
like using the  system and feel  confident about the numbers  they get
from neural networks. This is a often difficult task.

Prof. Grabec pointed  out  that the neural  networks, as  we know them
today,  were not very  suitable for  description  of and treatment  of
fields. 

Next  year,    we  also  intend    to  have a   session    on Japanese
state-of-the-art.   The  Japanese   have   been ahead   of others   in
applications of neural networks, and we hope  to be able to learn from
their experiences.