BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Talk of Professor Taiji Suzuki (University of Tokyo)
DTSTART;VALUE=DATE-TIME:20191101T120000
DTEND;VALUE=DATE-TIME:20191101T140000
UID:392cb142be70e58f85208ee50e36385e62ea16b8803571bfde3c7948
CATEGORIES:Conferences - Seminars
DESCRIPTION:Professor Taiji Suzuki\nTitle:\nGeneralization analysis and op
timization of deep learning: adaptivity and kernel view\n\nAbstract: In th
is talk\, I will discuss the adaptivity of deep learning\, and the general
ization ability and optimization property under overparameterized settings
. In the first half\, we theoretically show that deep learning can extract
appropriate bases in an adaptive way and thus can achieve superior perfor
mance than kernel methods especially on models with non-convexity property
. Thanks to this properties\, deep learning can outperform kernel methods
if input data are high dimensional and the target functions are in Besov s
pace.\nIn the later half\, we discuss the generalization ability and optim
ization property of deep learning under overparameterized settings. The cl
assical learning theory suggests that overparameterized models cause overf
itting. However\, practically used large deep models avoid overfitting\, w
hich is not well explained by the classical approaches. To resolve this is
sue\, we give a new unified frame-work for deriving a compression based bo
und. The existing compression based bounds can only be applied to a compre
ssed network\, but our bound can convert those bounds to that of non-compr
essed original network. Finally\, we discuss the optimization aspects of n
eural networks under the neural tangent kernel regime. We show that for a
classification task\, the width of networks can be much smaller than exist
ing studies to obtain a near global optimal solution by a gradient descent
.\n\nBIO: Taiji Suzuki is currently an Associate Professor in the Departme
nt of Mathematical Informatics at the University of Tokyo. He also serves
as the team leader of "deep learning theory group" in AIP-RIKEN. He receiv
ed his Ph.D. degree in information science and technology from the Univers
ity of Tokyo in 2009. He has a broad research interest in statistical lear
ning theory on deep learning\, kernel methods and sparse estimation\, and
stochastic optimization for large-scale machine learning problems. He serv
ed as technical program committee members of premier conferences such as N
eurIPS\, ICML\, ICLR\, COLT\, AISTATS and ACML. He received Outstanding Ac
hievement Award in 2017 from the Japan Statistical Society\, Outstanding A
chievement Award in 2016 from the Japan Society for Industrial and Applied
Mathematics\, and Best Paper Award in 2012 from IBISML.
LOCATION:INM 10 https://plan.epfl.ch/?room==INM%2010
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR