BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:UW-Madison-Physics-Events
BEGIN:VEVENT
SEQUENCE:2
UID:UW-Physics-Event-6872
DTSTART:20220316T160000Z
DTEND:20220316T171500Z
DTSTAMP:20260414T153725Z
LAST-MODIFIED:20220315T163427Z
LOCATION:Zoom link: https://lmu-munich.zoom.us/j/98576256641?pwd=QlVTW
 kxCclBIV2dOUkFOdXQ4ZWNEdz09#success
SUMMARY:BI for AI: Energy conserving descent for optimization\, Physic
 s ∩ ML Seminar\, Eva Silverstein\, Stanford University
DESCRIPTION:We introduce a novel framework for optimization based on e
 nergy-conserving Hamiltonian dynamics in a strongly mixing (chaotic) r
 egime and establish its key properties analytically and numerically. T
 he prototype is a discretization of Born-Infeld dynamics\, with a squa
 red relativistic speed limit depending on the objective function. This
  class of frictionless\, energy-conserving optimizers proceeds unobstr
 ucted until slowing naturally near the minimal loss\, which dominates 
 the phase space volume of the system. Building from studies of chaotic
  systems such as dynamical billiards\, we formulate a specific algorit
 hm with good performance on machine learning and PDE-solving tasks (so
  far at small scale)\, including generalization. It cannot stop at a h
 igh local minimum and cannot overshoot the global minimum\, proceeds f
 aster than GD+momentum in shallow valleys\, and predictably finds mult
 iple solutions according to a concrete formula for the measure on phas
 e space which is applicable as a result of the energy conservation. La
 rger-scale experiments in progress are required to assess its relative
  performance on ML problems of current interest\, along with further t
 heoretical analysis its impact on representation/feature learning. Bas
 ed on https://arxiv.org/abs/2201.11137 and ongoing work.
URL:https://www.physics.wisc.edu/events/?id=6872
END:VEVENT
END:VCALENDAR