BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:IC Colloquium : Probabilistic representation learning and scalable
Bayesian inference
DTSTART;VALUE=DATE-TIME:20171012T161500
DTEND;VALUE=DATE-TIME:20171012T173000
UID:db3bfdce0aa59704867b0c4a4deed50145b6fd81c1211d82778435a8
CATEGORIES:Conferences - Seminars
DESCRIPTION:**By : **Stephan Mandt - Disney Research Pittsbur
gh

\n

\n**Abstract :** Probabilistic modeling is a powe
rful paradigm which has seen dramatic innovations in recent years. These i
nnovations in approximate inference\, mainly due to automatic differentiat
ion and stochastic optimization\, have made probabilistic modeling scalabl
e and broadly applicable to many complex model classes. I start my talk by
reviewing the dynamic skip-gram model (ICML 2017) as an example of this c
lass. The model results from combining a probabilistic interpretation of w
ord embeddings with latent diffusion priors\, and allows us to study the d
ynamics of word embeddings for text data that are associated with differen
t time stamps. Our Bayesian approach allows us to share information across
the time domain\, and is robust even when the data at individual points i
n time is small. As a result\, we can automatically detect words that chan
ge their meanings even in moderately-sized corpora. Yet\, the model is Bay
esian non-conjugate\, and therefore we have to draw on modern variational
inference methods to train it efficiently on large data. The second part o
f my talk is therefore devoted to advances in variational inference. Here\
, I will review our very recent perturbative black box variational inferen
ce algorithm (NIPS 2017)\, that uses variational perturbation theory of st
atistical physics to construct corrections to the standard variational low
er bound. Last\, I will demonstrate that simple stochastic gradient descen
t with a constant step size is a form of approximate Bayesian inference (J
MLR and ICML 2016).

\n

\n**Bio :** Stephan Mandt is a
Research Scientist and head of the statistical machine learning group at
Disney Research Pittsburgh\, co-located with CMU. Previously\, he was a po
stdoctoral researcher with David Blei at Columbia University (2014-2016)\,
and a PCCM postdoctoral fellow at Princeton University (2012-2014). Steph
an Mandt holds a Ph.D. in theoretical physics from the University of Colog
ne\, supported by the German National Merit Foundation. His interests incl
ude large-scale probabilistic modeling\, representation learning\, variati
onal inference\, and media analytics.

\n

\n**More information**
LOCATION:BC 420 https://plan.epfl.ch/?room=BC420
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR