BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Inference Group
SUMMARY:Differential Privacy and Probabilistic Inference -
Oliver Williams (Microsoft Research)
DTSTART;TZID=Europe/London:20100621T110000
DTEND;TZID=Europe/London:20100621T120000
UID:TALK25096AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/25096
DESCRIPTION:Differential privacy is a recent privacy definitio
n that permits only indirect observation of data h
eld in a database through noisy measurement. I wil
l show that there is a strong connection between d
ifferential privacy and probabilistic inference. P
revious research on learning from data protected b
y differential privacy has been driven by research
ers inventing sophisticated learning algorithms wh
ich are applied directly to the data and output mo
del parameters which can be proven to\nrespect the
privacy of the data set. Proving these privacy pr
operties requires an intricate analysis of each al
gorithm on a case-by-case basis. While this does r
esult in many valuable algorithms and results\, it
is not a scalable solution for two reasons: first
\, to solve a new learning problem\, one must inve
nt and analyze a new privacy-preserving algorithm\
; second\, one must then convince the owner of the
data to run this algorithm. Both of these steps a
re challenging. In contrast\, I will consider the
potential of applying probabilistic inference to
the measurements and measurement process to derive
posterior distributions over the data sets and mo
del parameters thereof\, showing that for the mode
ls investigated\, probabilistic inference can impr
ove accuracy\, integrate multiple observations\, m
easure uncertainty\, and even provide posterior di
stributions over quantities that were not directly
measured.
LOCATION:TCM Seminar Room\, Cavendish Laboratory\, Departme
nt of Physics
CONTACT:Emli-Mari Nel
END:VEVENT
END:VCALENDAR