BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:An Algorithmic Perspective on Realism and Perceptual Quality in Lo
 ssy Compression - Yassine Hamdi (Imperial College London)
DTSTART:20250807T100000Z
DTEND:20250807T110000Z
UID:TALK235009@talks.cam.ac.uk
DESCRIPTION:Realism constraints (or constraints on perceptual quality) hav
 e received considerable recent attention in a diverse range of machine lea
 rning problems\, including within the context of lossy compression\, parti
 cularly of images. The most common recent theoretical formulation of reali
 sm\, inspired from generative modelling theory\, is distribution matching\
 , and involves a divergence between the distribution of generated samples 
 and some target distribution. More recently\, Theis argued that realism is
  best formalized via the algorithmic information theory of randomness test
 s\, particularly universal tests\, and that the latter are adequate for va
 rious machine learning problems involving realism\, such as outlier detect
 ion and generative modelling. In this talk\, we present theoretical eviden
 ce in support of this claim. We shall elaborate on the adequacy of univers
 al tests in the context of lossy compression. Theoretical studies of lossy
  compression with distribution matching constraints indicate that high-rat
 e common randomness between the compressor and the decompressor is a valua
 ble resource for achieving realism. On the other hand\, the utility of sig
 nificant amounts of common randomness has not been noted in practice. We o
 ffer an explanation for this discrepancy by considering a realism constrai
 nt that requires satisfying a universal test that inspects realizations of
  individual compressed reconstructions\, or batches thereof. Our results a
 lso provide a new perspective on the distribution matching formalism.
LOCATION:Seminar Room 2\, Newton Institute
END:VEVENT
END:VCALENDAR