Visual inference with hierarchical birth-death dynamics

  • Datum: 14.11.2019
  • Uhrzeit: 11:15 - 12:15
  • Vortragende(r): Jochen Braun
  • Otto-von-Guericke University Magdeburg
  • Ort: MRZ
  • Raum: Seminar Room
  • Gastgeber: Anna Levina
Visual inference with hierarchical birth-death dynamics

Sensory inputs are volatile and unpredictable and a reliable inference of underlying causes is difficult. Neural activity continuously accumulates and compares sensory evidence as perceptual decisions are reached. Psychophysical evidence from multistable perception, properly interpreted, elucidates many details of this stochastic dynamics. I show that the invariant characteristics of multistable perception are reproduced quantitatively by a competitive hierarchy of ‘birth-death' processes mimicking the metastable attractor dynamics recently postulated for cortical networks. The lower level of this hierarchy incrementally accumulates visual evidence, while the upper level categorically represents perceptual choice and phenomenal appearance, and also resets the lower level in preparation for future choices. I further show that this kind or representation provides clear functional benefits for volatile and unpredictable signals. Astonishingly, summing the birth-death representations of sensory signals can represent their shared mean more accurately than summing the signals directly (i.e., more accurately than the maximum likelihood estimate!). I conclude that a hierarchical birth-death dynamics has surprising explanatory power at several levels: dynamics of multistable perception, enhancing sensory function in a volatile and unpredictable world, and the 'ongoing' intrinsic dynamics of cortical activity.

Cao, Pastukhov, Mattia, Braun (2016) Collective activity of many bistable assemblies reproduces characteristic dynamics of multistable perception. J. Neurosci.,36: 6957-72.

Veliz-Cuba, Kilpatrick, Josic (2015) Stochastic models of evidence accumulation in changing environments. SIAM Review, 58: 264-89.

Zur Redakteursansicht