[Ag-chernov] [Ins-mitarbeiter] Vortrag Do 14 c.t.
Liebe Kollegen und Mitarbeiter, ich habe am Donnerstag Herrn Dr. Philipp Hennig vom MPI fuer Intelligente Systeme/Tuebingen zu Gast. Ich moechte Sie zu seinem Vortrag am Donnerstag, den 4.4. um 14 Uhr c.t. im Seminarraum 6.010, Wegelerstr. 6, einladen. Viele Gruesse, Carsten Burstedde Title: The Probabilistic View on Quasi-Newton Methods Quasi-Newton methods are a popular family of algorithms for continuous nonlinear optimization (one member, the BFGS method, might be the most widely used approach for this setting). Starting from a broader perspective on the connections between numerical methods and probabilistic inference, I will show that quasi-Newton algorithms can be interpreted as least-squares/Gaussian regressors, establishing a connection between optimization and regression. Among the ideas arising immediately from this is a nonparametric generalisation of quasi-Newton methods to a Gaussian-process prior. This extension does not affect computational complexity, but increases the reach of quasi-Newton methods in various ways I will touch upon, including optimization of noisy and dynamically changing objectives. It also poses new analytical challenges yet to be addressed. Most of the talk will proceed on a comparably conceptual level, only superficial knowledge in optimization and probabilistic inference is required. _______________________________________________ Ins-mitarbeiter mailing list Ins-mitarbeiter@ins.uni-bonn.de https://mail.ins.uni-bonn.de/mailman/listinfo/ins-mitarbeiter
Hier noch eine Erinnerung an den Vortrag: Insbesondere interessant in Bezug auf Optimierung, inverse Probleme, Parameterschaetzung und UQ. Viele Gruesse, Carsten On 04/02/2013 02:24 PM, Carsten Burstedde wrote:
Liebe Kollegen und Mitarbeiter,
ich habe am Donnerstag Herrn Dr. Philipp Hennig vom MPI fuer Intelligente Systeme/Tuebingen zu Gast. Ich moechte Sie zu seinem Vortrag am Donnerstag, den 4.4. um 14 Uhr c.t. im Seminarraum 6.010, Wegelerstr. 6, einladen.
Viele Gruesse,
Carsten Burstedde
Title: The Probabilistic View on Quasi-Newton Methods
Quasi-Newton methods are a popular family of algorithms for continuous nonlinear optimization (one member, the BFGS method, might be the most widely used approach for this setting). Starting from a broader perspective on the connections between numerical methods and probabilistic inference, I will show that quasi-Newton algorithms can be interpreted as least-squares/Gaussian regressors, establishing a connection between optimization and regression. Among the ideas arising immediately from this is a nonparametric generalisation of quasi-Newton methods to a Gaussian-process prior. This extension does not affect computational complexity, but increases the reach of quasi-Newton methods in various ways I will touch upon, including optimization of noisy and dynamically changing objectives. It also poses new analytical challenges yet to be addressed.
Most of the talk will proceed on a comparably conceptual level, only superficial knowledge in optimization and probabilistic inference is required.
Ins-mitarbeiter mailing list Ins-mitarbeiter@ins.uni-bonn.de https://mail.ins.uni-bonn.de/mailman/listinfo/ins-mitarbeiter
participants (1)
-
Carsten Burstedde