
I
suppose it is about
time to update my old web page and put on it all kinds of stuff that
happened since I last touched HTML code 6 years ago. So, anyway, it has
been 7 years since I received the little yellow slip of paper
indicating that all requirements set forth by MIT for granting me a
degree of Doctor of Philosophy (I still had to look up how to spell it)
have been satisfied and I am now free to pursue my other interests. It
was another way of saying that they will no longer pay my tuition and
would I be so kind to get the hell out of the office provided to me by
the MIT corporation.
Recent Projects
A
large undertaking
with
Chris
Wren at
MERL.
The project is based on the idea that general rough context of the
human environment can be glipmsed from a large array of low fidelity
sensors. This context can provide a set of powerful cues about human
behavior, be it an individual, or a large organization. This project
has deep implications for social studies, security and building
management applications. And is just an interesting thought that went
way too far. I am glad that Chris and I managed to stick with it for
this long - it got quite a bit of academic attention.
<more to come>
Category learning with weakly labeled
data
This
was the topic of my dissertation. This idea develops some esoteric but
practical aspects of machine learning that explain the process of
long-term category forming.
This is
Bill
Tomlinson's baby, to which a few of us contributed. At this
time some of my on-line audio learning code got into this system but
most of my time was really taken with writing my dissertation. Bill was
fascinated with the development of social structure in a pack of wolves
and built this system to try to mimic it in a realistic way. I love the
visuals, due to
Marc
Downie and
Adolph
Wong's fascinating animations.
This was the first project that I worked on withe the (void *) crew
headed by
Bruce
Blumberg.
The project is devoted to trying to understand high level learning
processes that take place during animal training. Domesticating of
animals has been going on for thousands of years and animal trainers
have picked up some tricks along the way that work pretty much as seen
on TV. We were curious about what it would take to build machines that
can be trained to be as obedient as dogs without involving linguistics,
speech and gesture recognition. It clearly can be done in principle,
since dogs do not follow the language development paths of humans, yet
are perfectly capable of responding to simple voice commands. This
project approached the problem of learning grounded speech using
actions as labels for "what matters" to an animal about human speech.
Scientific American Frontiers had a
nice
episode about it.
KidsRoom
was a one-of-a-kind
application of computer vision for a physically situated
computer
game. It took a few grad students from Aaron Bobick's
group (back when he was at MIT) about 6 weeks to build and was
only
shown
publicly twice. However, its distant cousin, "KidsRoom 2",
was re-made commercially by a local company "NearLife"
and was exhibited for a few months in the London Millenium Dome in
2000. Here are some
videos kindly digitized by Claudio
Pinhanez with my comments.
Send me mail at yivanov at gmail.com.