BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Charles O'Neill & Jack Miller (ANU)
DTSTART:20220823T060000Z
DTEND:20220823T070000Z
DTSTAMP:20260423T005820Z
UID:anumacs/12
DESCRIPTION:Title: <a href="https://researchseminars.org/talk/anumacs/12/"
 >Eigenvalue initialisation and regularisation for koopman autoencoders and
  beyond</a>\nby Charles O'Neill & Jack Miller (ANU) as part of ANU Mathema
 tics and Computational Sciences Seminar\n\nLecture held in Room 1.33\, Han
 na Neumann Building #145.\n\nAbstract\nRecent efforts have been made to le
 arn the Koopman operator with predictive autoencoders. However\, little at
 tention has been payed to the initialisation of these networks. Noting the
  importance of eigenvalues to the action of a linear operator\, one may as
 k whether it would be useful to employ them in the initialisation and regu
 larisation of these autoencoders? To answer this\, we devise a spectral ei
 genvalue initialisation and eigenvalue penalty scheme. Having done so\, we
  discover that eigenvalues do in fact have great utility for this purpose.
  We demonstrate that in learning a Koopman operator for a damped driven pe
 ndulum\, appropriate initialisation and regularisation can improve initial
  performance by an order of magnitude. We also show with this system that 
 as the dissipative element of a dynamical system decreases\, the utility o
 f unit circle initialisation schemes increase and the utility of different
  regularisation schemes change. Additionally\, we show that the benefits o
 f eigenvalue initialisation and regularisation generalise to real-world cy
 clone wind data\, sea surface temperature prediction and flow over a cylin
 der.\n
LOCATION:https://researchseminars.org/talk/anumacs/12/
END:VEVENT
END:VCALENDAR