BEGIN:VCALENDAR
PRODID:-//Columba Systems Ltd//NONSGML CPNG/SpringViewer/ICal Output/3.3-
 M3//EN
VERSION:2.0
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
DTSTAMP:20251110T144626Z
DTSTART:20251029T130000Z
DTEND:20251029T140000Z
SUMMARY:SQUIDS Seminar (ProbAI-sponsored): Adaptive friction and nonlinea
 r damping for model training
UID:{http://www.columbasystems.com/customers/uom/gpp/eventid/}p16j-mht94l
 sd-umnhd1
DESCRIPTION:Speaker: Katerina Karoni (Bristol)\nAbstract: We discuss nove
 l damping procedures for training large scale Bayesian data models\, suc
 h as deep neural networks. Drawing inspiration from the concept of a the
 rmostat (widely used for temperature regulation in molecular dynamics)\,
  we introduce kinetic energy controls on the individual parameter veloci
 ties of the model. This approach can be likened to component-wise Nosé-H
 oover style thermostatting taken in the zero-temperature limit and it ca
 n be directly related to the introduction of cubic damping\, a vibration
  suppression mechanism used in structural engineering applications. Whil
 e a large momentum parameter helps to overcome barriers and progress tra
 ining in low-curvature regions\, it should be reduced in areas with stee
 p gradients to avoid instability\; our adaptive scheme allows this adjus
 tment to be performed automatically\, on a per-parameter basis. By using
  these schemes\, we obtain enhanced efficiency\, including significant s
 peedups and test accuracy improvements in representative deep learning t
 asks.
STATUS:TENTATIVE
TRANSP:TRANSPARENT
CLASS:PUBLIC
LOCATION:Frank Adams Room 2\, Alan Turing Building\, Manchester
END:VEVENT
END:VCALENDAR