WHAT
WHY
HOW
Bayesian InferenceBayesian Inference
Flip a coin.
Event: A, B
A: One head at least
B: Both heads or tails
Question:
P(B|A)?
He didn’t say “hi” when passing by.
Compare P(h1|D) and P(h2|D) to reach the most likely assumption.
P(h2|D)=P(h2)*P(D|h2)/P(D)
P(h1|D)=P(h1)*P(D|h1)/P(D)
P(Bi|A) P(Bi) * P(A|Bi) ∝
P(hi|D) P(h∝ i) * P(D|hi)
priori probability
WHAT
WHY
HOW
Bayesian InferenceBayesian Inference
• Complexity• Ambiguity• Duplicity• Insufficiency
One object? Two objects?
What is perceived are essentially those objects and events that under normal conditions would be most likely to produce the received sensory stimulation.
What is perceived are essentially those objects and events that under normal conditions would be most likely to produce the received sensory stimulation.
WHAT
WHY
HOW
Bayesian InferenceBayesian Inference
P(h|D) P(h) * P(D|h) ∝
‘‘Calling some skill or behavior innate tends to stop analysis of how it
develops’’ (Fischer & Stewart, 1999, p. 150).
WRONG!
Ina
Innate assumptions validated by comparative research (Hershberger, 1970)
•Single light source.•Light source above.
Natural Image Parsing
N: non-terminal nods (models)
Evidence from behavioral research
Evidence from neural science• ERP and fMRI (Murray, et al, 2002): Dorsal regions provide input to lateral occipital complex (LOC), where
signatures of illusory contour completion first appear, then followed by activity in V1/V2.
• MEG and fMRI (Bar et al, 2006): Low spatial frequency specific activity appeared in left orbitofrontal
cortex 50 ms earlier than in fusiform gyrus of temporal cortex.
• Bayesian Ideal Observer
• And more…
Vision as Bayesian inference: analysis by synthesis? Alan Yuille and Daniel Kersten, 2006
ReferencesReferences
平凡而又神奇的贝叶斯方法刘未鹏 , 2008
Innateness (and Bayesian) Visual PerceptionBrian Scholl, 2005