Reminder : this happens in 1h !
Le 11 déc. 2018 à 17:20, rim.assouel@gmail.com a écrit :
This week we have Praneeth from MSR India giving a talk on Thursday December 13 2018 at 15:30 in room AA3195
Will this talk be streamed https://mila.bluejeans.com/4255239897/webrtc? Yes
Pay attention to the day change !! This tea talk will happen on THURSDAY :)
As it will be the last tea talk of the year, it will be featured with actual tea, talk and snacks :)
See you there! Rim and Sai
TITLE On momentum methods and acceleration in stochastic optimization
ABSTRACT
It is well known that momentum gradient methods (e.g., Polyak's heavy ball, Nesterov's acceleration) yield significant improvements over vanilla gradient descent in deterministic optimization (i.e., where we have access to exact gradient of the function to be minimized). However, there is widespread sentiment that these momentum methods are not effective for the purposes of stochastic optimization due to their instability and error accumulation. Numerous works have attempted to quantify these instabilities in the face of either statistical or non-statistical errors (Paige, 1971; Proakis, 1974; Polyak, 1987; Greenbaum, 1989; Roy and Shynk, 1990; Sharma et al., 1998; d’Aspremont, 2008; Devolder et al., 2013, 2014; Yuan et al., 2016) but a precise understanding is lacking. This work considers these issues for the special case of stochastic approximation for the linear least squares regression problem, and shows that:
- classical momentum methods (heavy ball and Nesterov's acceleration) indeed do not offer any improvement over stochastic gradient descent, and
- introduces an accelerated stochatic gradient method that provably achieves the minimax optimal statistical risk faster than stochastic gradient descent (and classical momentum methods).
Critical to the analysis is a sharp characterization of accelerated stochastic gradient descent as a stochastic process. While the results are rigorously established for the special case of linear least squares regression, experiments suggest that the conclusions hold for the training of deep neural networks. BIO Praneeth Netrapalli is a researcher at Microsoft Research India, Bengaluru since August 2016. Prior to this, he was a postdoctoral researcher at Microsoft Research New England in Cambridge, MA. He obtained MS and PhD from UT Austin and B-Tech from IIT Bombay all in Electrical Engineering. His research focuses on designing efficient algorithms for machine learning problems primarily via stochastic and nonconvex optimization. More information about his research is available on his home page http://praneethnetrapalli.org/