Illuminating the hierarchical dynamics of visual inference

Friday, 2017, March 31 - 12:00
Cognitive Biology Group, OVGU, Magdeburg, Germany


The visual environment is noisy, variable, and ambiguous.  In the face of unknown prior statistics, optimal inference is computationally challenging even in simplified settings.  However, optimal inference can be approximated by accumulating and comparing evidence in stochastic dynamical systems, which also suggest plausible neural implementations (Bogacz et al., 2006; Deneve, 2007; Veliz-Cuba, Kirkpatrick, Josic 2016).   But what kind of stochastic dynamics governs visual perception in human observers and how suitable is this dynamics for performing inference?  
The perceptual dynamics engendered by multi-stable visual displays appears to comprehensively illuminate these issues.  Specifically, some well-established and generic characteristics -- abrupt onset of perception & gradual adaptation/recovery of input representations, stereotypical shape of dominance distribution (“scaling property”), bimodal dependence on input strengths (“Levelt’s propositions”) -- suffice to fully constrain an underlying stochastic dynamics with (at least) three hierarchical levels: bistable units biased by stimulation (cortical assemblies? cortical columns?), uncoupled populations of units to accumulate evidence, coupled populations to represent perception, and quantified, reciprocal interactions between all levels.    
With suitable input weights, this empirically identified dynamical system accumulates noisy evidence to make nearly optimal categorical discriminations.  The identified dynamics has several unexpected features, such as discrete stochasticity and accelerated discounting of evidence already perceived.  These features seem to offer important functional benefits in a volatile worlds, such as fast convergence of inference, or balancing stability and sensitivity of inference.
Suggested Readings:
Cao, Pastukhov, Mattia, Braun (2016) Collective activity of many bistable assemblies reproduces characteristic dynamics of multistable perception.  J. Neurosci., 36: 6957-72. 1.pdf
Cao, Braun, Mattia (2014) Stochastic accumulation by cortical columns may explain the scalar property of multistable perception.  Phys. Rev. Lett., 113: 098103 2.pdf

Further Readings:

Pastukhov, Garcia-Rodriguez, Haenicke, Guillamon, Deco, Braun (2013)  Multi-stable perception balances stability and sensitivity. Front. Comput. Neurosci., 7: 17.
Braun, Mattia (2010) Attractors and noise: twin drivers of decisions and multistability.  Neuroimage, 52: 740-51.
Bogacz et al., (2006) The physics of optimal decision making: a formal analysis of models of performance in two-alternative forced-choice tasks.  Psychol. Rev., 113: 700-765.
Deneve (2008) Bayesian spiking neurons I: Inference.  Neural Computation, 20, 91-117.
Veliz-Cuba, Kilpatrick, Josic (2016) Stochastic models of evidence accumulation in changing environments.  SIAM Review, 58: 264-289.