13 / 24
Elisa explains that the main problem
in domain adaptation is how to
measure domain discrepancy and how
to cope with it. There are many
strategies for
deep domain
adaptation, but the question is: which
one is the best? Their strategy of
adaptive batch normalisation has
proved to be very effective, but they
know that there are others.
The idea is to adapt the batch
normalisation layers, that are very
popular layers in convolutional neural
networks, and use different statistics
for the source and target domain.
Elisa says: “I think one interesting
thing about our approach is that it is
the first method where at each layer
of the network,
the network
automatically chooses how much to
adapt. This, to my knowledge, has not
been done in previous work and I
think it is one of the strengths of our
approach. This leads to very good
experimental results.”
One feature the team would like to
add to the method is a way to better
measure the distance between the
distributions. They are using a simple
way to model
distributions,
considering just the mean and
variance, but Elisa says they could
explore using high-order statistics,
histogram matching, or another more
sophisticated way to change these
layers. They want to keep the idea of
automatically choosing at each level
of the network how much to adapt.
As for next steps, Elisa tells us that
there is more to investigate: “
This
work can be extended to a wide range
of applications. Now, we use a
common benchmark based on object
recognition, but I think the work can
be extended to many other problems.
There are a lot of problems where you
need to cope with domain shift. For
instance,
you have different
illumination conditions, you have
people moving from different
cameras. Investigating the applicative
point of view, it is important. We
should move there
.”
If you want to learn more about
this work, visit Elisa’s poster
today (Friday) 10:30 at Sala
Mosaici 2.
13
Friday
Elisa Ricci
“
The main problem in domain adaptation
is how to measure domain discrepancy
and how to cope with it
”
“
The work can be extended to many other
problems. There are a lot of problems
where you need to cope with domain shift
”