Perceptual Multistability as Markov Chain Monte Carlo Inference.
-
Upload
karin-marshall -
Category
Documents
-
view
216 -
download
0
Transcript of Perceptual Multistability as Markov Chain Monte Carlo Inference.
3
Outline
Construct rational model of visual process.• Algorithmic/computational model of mental
processing– Not about neurons– Bayesian inference promising, computationally infeasible
• Explain existing results– Multistability– Focus on binocular rivalry
4
Your brain is flat out making stuff up
Sensory inputs fundamentally impoverished.• Reconstructing 3D world from 2D vision
Bayesian inference promising• Belief (posterior) computed from sensory inputs
(likelihood) and plausible world structures (prior)• Effective in practice• Requires approximation• Prior proposed approximation don’t represent
uncertainty
7
Your brain is flat out making stuff up
Sensory inputs fundamentally impoverished.• Reconstructing 3D world from 2D vision
Bayesian inference promising• Belief (posterior) computed from sensory inputs
(likelihood) and plausible world structures (prior)• Effective in practice• Requires approximation• Prior proposed approximation don’t represent
uncertainty
8
Your brain knows it’s making stuff up
Sensory inputs fundamentally impoverished.• Reconstructing 3D world from 2D vision
Bayesian inference promising• Belief (posterior) computed from sensory inputs
(likelihood) and plausible world structures (prior)• Effective in practice• Requires approximation• Prior proposed approximation don’t account for
multistability
9
Outline
New model of approximation (MCMC)• Exploration of hypothesis space corresponds to
switches in multistability.• MCMC gives accurate predictions of state
distributions!
11
Modelling the visual process
Latent imageOutlier process
Retinal stimulus
Relationship among pixels:
17
Summary
Rational model of visual process multistability• Simple model of visual process (Bayesian)• Standard approximation techniques from machine
learning (MCMC)• Accurately predicts experimental results, including
multistability in binocular rivalry• Provides high-level intuition of neurally-plausible
explanations.