so I have been doing deep learning
maybe over about 10 years I mean I did
PhD at Stanford under professor andry so
I was pretty much the first student in
his group starting doing deep learning
for PhD and also it was the beginning
period where other labs had some so like
very emerging research on deep learning
so so since then I got PhD in 2010 and
then moved to Michigan as a seasoned
professor
so my work is about learning a
representation that can find some
invariant representation that can do
well on many different classification
tasks and also more recently I've been
working on representation on skin that
can be useful for doing inference and
reasoning and in terms of application I
have been working on computer vision
problems quite extensively ranging many
different type of questions such as
image classification object detection
segmentation and so on I also work on a
problem called multimodal learning which
means that we want to find some
high-level deep representation that can
connect many different heterogeneous
types of input modalities such as images
and text or audio and video and also
other time series data so how to find a
good share representation among these
heterogeneous input data is one of the
research team I have been working on
so at that time there were some interest
in how to build or some algorithm that
behave like some brain I mean so there
was a very high-level motivation but
I've been always interested in building
some general-purpose algorithms that can
somehow learn everything from scratch so
I was influenced by people like Bruno
thousand and Tommy Poggio and they
developed models such as sparse coding
and hierarchical necks that kind of
explains the early visual cortex so
there has been some evidence that if you
do some learning and build some
hierarchical structure you can learn
some interesting features that it can
explain the phenomena that happens in
the brain so there was one motivation
but in general I want also have been
very interested in how to build a
general-purpose machinery that can just
learn from the data without too much
engineering of the like low-level
features yeah so I feel that one big
application of the medical domain so
building some smart algorithms that can
assist on doctors in diagnosis and other
decision making will be extremely
helpful and also lots of hospitals and
companies and other research labs are
interested in building the deep learning
systems for medical applications so it
looks very promising to me another
interesting very exciting application
area is some kind of building on agent
so personal assistant agent that can
help humans to
do some some intelligent tasks and save
time and maybe give some advice so there
are there are many different application
areas that can be built but this will
require some maybe new class of learning
algorithms that can interact with humans
and also make some sequential decision
making so in that context maybe more
development in reinforcement learning or
our so-called deep learning
reinforcement learning will be important
and also one challenge is how to deal
with a small number of small amount of
label data so so how to learn a good
system without having to rely on a huge
number of manual label there will be
also important challenge and especially
tackling some important problems in
medicine so yeah one one question is
really about this how to start with like
very little prior knowledge or how to
start with not so many labels and also
how do we actually make the learning
algorithm to somehow interact with the
environment and learn from interaction
so one motivating example is robots so
robots would actually I mean ideally I
mean you can think about humans there's
also some sort of Asian who actually can
learn by interacting with the
environment or interacting with the
object so potentially if we have really
great progress in robotics and deep
learning it's part it may be possible
that robots can also just start
exploring the world and then learn
something useful out of just doing this
interaction so in order to solve this
problem I think that so there's a lot of
challenges in solving perception and
also decision-making control and all
things putting together so
is those challenging problem but at the
same time it sounds very exciting to me
I'm personally most interested in
combining deep learning and
reinforcement learning because I believe
that building some agent that can
interact and perform actions in the real
world is very very important and also it
will bring much more value than just
doing perception so for example just
recognizing images it's useful but in
order to come in build some high value
activities
it's much more useful if the agent can
perform some action on top of the
perception so in order to solve this
problem I think reinforcement learning
is one critical bottleneck of this
problem so I believe that the fifth
reinforcement learning is one of the
most exciting and frontiers of deep
learning yeah yeah I loved just
listening to many broad like broad range
of speakers who are some of our from
academia and some of our from startups
especially I'm interested in how deep
learning can be used in useful in many
some industrial applications and also
emerging topics in startups so it was a
very interesting and exciting topics and
also I liked about how to learn some
textual relationships from the using
deep learning algorithms so it was a
talk by Professor Andrew McCollum so I
enjoyed the talk particularly yeah
you
카테고리 없음