Artificial Neural Networks for Pattern Recognition:

Anything news under the stars?

Marco Gori and Simone Marinai

TC3 Chair and Vice Chair

In the last fifteen years, we have seen impressive advances in the field of

pattern recognition using connectionist models. The truly sub-symbolic nature of most

problems makes these models, which strongly emphasize learning, a valuable alternative

to related decision-theoretic approaches. Somebody argue that their biological inspiration

is the key of their inherent power, others simply emphasize classic methodological issues

and claim that they are the reason of most successful experimental results. Regardless

of the where the truth lies, many pattern recognition tasks have witnessed successes and

failures which have not been completely understood yet. Rigorous experimental analyses

are still required to draw a reasonable picture of the capabilities of nowadays neural

networks for pattern recognition tasks. On the other hand, somebody argues that new frontiers in the field of neural networks might open the doors to significant achievements.

Amongst recent developments, which range from strongly biologically inspired neural networks

to strongly mathematical based models, like kernel machines, the extension to learning

in structured domains seems to be one of the most promising research area in the next

few years for the application to pattern recognition tasks.

Classic neural networks-based approaches have been properly extended to

the case in which the patterns are given rich graphic representations.

This approach, which is referred to as adaptive

graphical pattern recognition, is somehow in between decision-theoretic and structural

pattern recognition, and it is very promising direction in the field of pattern recognition.

Interestingly enough, some classic algorithms like Backpropagation are extended in such

a way to deal with graphical inputs. The corresponding learning algorithms

turn out to be very related to graph matching and other methods used in structural pattern

recognition. This new methodology represents and interesting challenge for researchers

in the field of pattern recognition, since many typical tasks faced using feature-based

representations (vectors) are likely to be attacked from a completely different point

of view, in which the pattern structure is strongly emphasized. Interestingly enough,

these structural domains are not equipped with a special metrics but, on the opposite, similarities and pattern classification are performed on the basis of learning from examples. However, this is just a case of many different hybrid schemes which go beyond the

brute force of learning and try to incorporate prior knowledge as much as possible.

One of the tasks of TC3 (http://www.dii.unisi.it/TC3/) is to stimulate the discussion

on recent advances in the field of neural networks and pattern recognition so as to highlight most promising

research directions. We have seen massive efforts for the experimentation of most

traditional architecture and learning algorithms, like LVQ and Backpropagation, but,

as an Irish proverb puts it, "There are finer fish in the sea than have ever been caught."

We really hope that the TC3 will be able to offer some guidelines for discovering finer fish.

Novel approaches, like learning in structured domains and other hybrid approaches,

have only been preliminarily investigated. The first official opportunity for bringing together researchers in the field will be

the workshop "Artificial Neural Networks in Pattern Recognition", which will be held in

Florence (Italy) on 12-13 September 2003. There will be a strong emphasis on a contest, which will

play a central role in the workshop. We really hope that this first workshop will be

a lively market place for fine fish.