[YPNG] YPNG Friday 5 April 2019

[Sekhar Tatikonda]

Hi Folks,

Dominic Richards (visiting from Oxford) will talk this Friday about:

Title: Optimal Statistical Rates for Non-parametric Decentralised
Regression with
Distributed Gradient Descent

Abstract: Due to bandwidth limitations, privacy concerns or network
instability, it is
often required to fit statistical models on data sets stored across
multiple computers
without a central server to coordinate computation and disseminate
information i.e. star
topology. This has motivated the investigation of decentralised methods
which solve the
problem in a more robust manner by not relying on a single computer. In
this work we
investigate the statistical performance of a simple synchronous
decentralised iterative
gradient descent method (Distributed Gradient Descent) in the homogeneous
distributed
non-parametric regression setting i.e. computers hold samples from the same
distribution.
By utilising the concentration of quantities held by individual computers,
we show there
are a number of settings where computers can save on computational and
communication
costs without any loss in statistical accuracy. Given computers hold
sufficiently many
samples with respect to the network topology, we show that Distributed
Gradient Descent
yields optimal statistical rates with the same numbers of iterations as
Centralised algorithm.
(Joint work with P. Rebeschini)

See you Friday at 11am in the Stat's classroom.

Regards,
sekhar
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.yale.edu/pipermail/ypng/attachments/20190402/5b2b3e39/attachment-0001.html>