14 October 2015
San Francesco - Via della Quarquonia 1 (Classroom 1 )
Limitations of shallow (one-hidden-layer) feedforward networks to
represent efficiently finite mappings will be discussed. It will
be shown that almost any uniformly randomly chosen mapping on a
sufficiently large finite domain cannot be tractably represented
by a shallow network with perceptrons or kernel units. Existential
probabilistic result will be illustrated by a concrete example of
a class of such functions constructed using quasi-random sequences.
Analogies with central paradox of coding theory and the "no-free-lunch
theorem" will be discussed.
relatore:
Kurkova, Vera
Units:
DYSCO