La Sezione di Informatica 
del Dipartimento di Scienze Fisiche propone per giovedì 30 giugno 
due seminari:


Il Dott. Antonio ELEUTERI
Dip. Scienze Fisiche - Università di Napoli Federico II, 

giovedì 30 giugno alle 14:30

nell'aula 0M03 del DSF 
(Compl. universitario Monte S. Angelo - Via Cintia)

terrà il primo seminario, dal titolo: 
 

An information geometric approach to 
variable selection by neural networks

In questo seminario illustriamo un nuovo approccio, basato sulla 
geometria dell'informazione, per la selezione di variabili mediante reti 
del tipo Multi-layer Perceptron. Tale approccio è basato su proiezioni 
della varietà di Riemann definita da un MLP su sottovarietà definite da 
reti MLP con dimensione di input ridotta. Mostriamo come la divergenza 
informazionale tra modelli può essere usata come criterio per una 
ricerca efficiente nello spazio delle reti con differenti dimensioni 
dell'input. Inoltre, mostriamo come le probabilità a posteriori dei 
modelli possono essere valutate per ordinare i modelli proiettati. 
Infine, testiamo il nostro algoritmo su dati sintetici e reali, e 
confrontiamo le sue performance con altri metodi riportati in letteratura. 
 

Dott.ssa Claudia R. CALIDONNA 
Istituto di Cibernetica “E. Caianiello” del CNR (Arco Felice)

che, sempre giovedì 30 giugno, alle 16:30, nella stessa aula parlerà di 

Modeling and simulating 
by a hybrid cellular automata paradigm: 
implications and perspectives

Abstract:
The simulation of complex physical systems has always been a
challenging research area that involves different competencies and skills.
The cellular automata (CA) approach proved to be a simple mathematical model for
simulation application, such as statistical mechanics problems characterized by a number of
different physical parameters mutually coupled in a non-linear way Cellular automata
supply useful models for many investigations in natural science, combinatorial mathematics,
and computer science; in particular, they represent a natural way of studying the
evolution of large physical systems involving discrete co-ordinates and state
variables as well as discrete time steps The cellular automata computing model
looks like to provide new tools  for doing simulation more efficiently by offering
the potential of addressing issues as simplicity, locality, and parallelism. When successfully
addressed, these features can be exploited to improve cellular automata application performances.

As it is well known, the requirement of good performances is a very crucial
point in the simulation activity.

Another issue to be considered is that there are physical phenomena which cannot be expressed in
purely local terms as a function of the local cellular neighborhood, but
they may involve the computation of non local properties. In such cases it
is necessary to consider the whole grid in order to compute a correct output. Finding
local interaction rules to solve global problems is a considerable challenge for the
designer of the physical model.

Especially when addressing global problems, it could be difficult to
design highly local systems exhibiting a specific behavior: that implies the
extension of the CA classical paradigm through the introduction of new
characterizations and generalisations.

In this research activity an extended CA paradigm, the Cellular Automata
Network (CAN) was conceived: it introduces besides the inner data parallelism,
another source of parallelism (task parallelism) coming from the relation between the
Cellular Automata which belong to the network. CA networks can be executed
on a parallel computer.

Exploiting either the data parallelism or both control and data parallelism,
which results in a multi-level parallelism application, it is clear that policies
are necessary to decide the amount of parallelism to be spawned for each level
in order to have a better exploitation of the available parallel computer resources.

This work addresses the study and a further extension of the CAN paradigm, CAN v. 2,
allowing for the simulation of macroscopic phenomena by means of control constructs;
showing implications and perspectives involved with.

=================================================================================== 

il proponente

Prof. Ernesto Burattini