Title: Bayesian Perception
1Bayesian Perception
2General Idea
Ernst and Banks, Nature, 2002
3General Idea
Conditional Independence assumption
4General Idea
Generative model
w?
Ernst and Banks, Nature, 2002
5General Idea
Probability
Width
6General Idea
7General Idea
8General Idea
Probability
Width
v
t
9General Idea
10Optimal Variance
Fisher information sums for independent signals
11General Idea
Predicted by the Bayesian model
0.2
0.15
Threshold (STD)
0.1
0.05
0
0
67
133
200
Visual noise level ()
Note unimodal estimates may not be optimal but
the multimodal estimate is optimal
Ernst and Banks, Nature, 2002
12Adaptive Cue Integration
- Note the reliability of the cue change on every
trial - This implies that the weights of the linear
combination have to be changed on every trial! - Or do they?
13General Idea
- Perception is a statistical inference
- The brain stores knowledge about P(I,V) where I
is the set of natural images, and V are the
perceptual variables (color, motion, object
identity) - Given an image, the brain computes P(VI)
14General Idea
- Decisions are made by collapsing the distribution
onto a single value - or
15Key Ideas
- The nervous systems represents probability
distributions. i.e., it represents the
uncertainty inherent to all stimuli. - The nervous system stores generative models, or
forward models, of the world (P(IV)), and prior
knowlege about the state of the world (P(V)) - Biological neural networks can perform complex
statistical inferences.
16Motion Perception
17The Aperture Problem
18The Aperture Problem
19The Aperture Problem
20The Aperture Problem
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
21The Aperture Problem
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
22The Aperture Problem
23The Aperture Problem
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
24The Aperture Problem
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
25The Aperture Problem
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
26Standard Models of Motion Perception
- IOC interception of constraints
- VA Vector average
- Feature tracking
27Standard Models of Motion Perception
IOC
VA
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
28Standard Models of Motion Perception
IOC
VA
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
29Standard Models of Motion Perception
IOC
VA
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
30Standard Models of Motion Perception
IOC
VA
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
31Standard Models of Motion Perception
- Problem perceived motion is close to either IOC
or VA depending on stimulus duration,
eccentricity, contrast and other factors.
32Standard Models of Motion Perception
Percept VA
Percept IOC
IOC
IOC
VA
VA
Vertical velocity (deg/s)
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
Horizontal velocity (deg/s)
33Moving Rhombus
34Bayesian Model of Motion Perception
- Perceived motion correspond to the MAP estimate
35Prior
- Human observers favor slow motions
36Likelihood
37Likelihood
38Likelihood
Binary maskPresumably, this is set by
segmentation cues
39Posterior
40Bayesian Model of Motion Perception
- Perceived motion corresponds to the MAP estimate
Only one free parameter
41Likelihood
42Motion through an Aperture
- Humans perceive the slowest motion.
- More generally we tend to perceive the most
likely interpretation of an image
43Motion through an Aperture
Likelihood
50
Vertical Velocity
0
-50
-50
0
50
ML
Horizontal Velocity
50
50
Vertical Velocity
Vertical Velocity
MAP
0
0
-50
-50
Prior
Posterior
-50
0
50
-50
0
50
Horizontal Velocity
Horizontal Velocity
44Motion and Constrast
- Humans tend to underestimate velocity in low
contrast situations
45Motion and Contrast
Likelihood
50
Vertical Velocity
0
-50
High Contrast
-50
0
50
ML
Horizontal Velocity
50
50
Vertical Velocity
Vertical Velocity
MAP
0
0
-50
-50
Prior
Posterior
-50
0
50
-50
0
50
Horizontal Velocity
Horizontal Velocity
46Motion and Contrast
Likelihood
50
Vertical Velocity
0
-50
Low Contrast
-50
0
50
ML
Horizontal Velocity
MAP
50
50
Vertical Velocity
Vertical Velocity
0
0
-50
-50
Prior
Posterior
-50
0
50
-50
0
50
Horizontal Velocity
Horizontal Velocity
47Motion and Contrast
- Driving in the fog in low contrast situations,
the prior dominates
48Moving Rhombus
Likelihood
50
50
Vertical Velocity
Vertical Velocity
0
0
-50
-50
High Contrast
-50
0
50
-50
0
50
IOC
Horizontal Velocity
Horizontal Velocity
MAP
50
50
Vertical Velocity
Vertical Velocity
0
0
-50
-50
-50
0
50
-50
0
50
Prior
Posterior
Horizontal Velocity
Horizontal Velocity
49Moving Rhombus
Likelihood
50
50
0
Vertical Velocity
0
Vertical Velocity
-50
-50
Low Contrast
-50
0
50
-50
0
50
IOC
Horizontal Velocity
Horizontal Velocity
50
50
MAP
Vertical Velocity
Vertical Velocity
0
0
-50
-50
-50
0
50
-50
0
50
Prior
Posterior
Horizontal Velocity
Horizontal Velocity
50Moving Rhombus
51Moving Rhombus
Percept VA
Percept IOC
IOC
IOC
VA
VA
Vertical velocity (deg/s)
Vertical velocity (deg/s)
Horizontal velocity (deg/s)
Horizontal velocity (deg/s)
52Barberpole Illusion
53Plaid Motion Type I and II
54Plaids and Contrast
Lower contrast
55Plaids and Time
- Viewing time reduces uncertainty
56Ellipses
57Ellipses
- Fat vs narrow ellipses
- All motions agree
58Ellipses
59Ellipses
- Adding unambiguous motion
60Ellipses
- Adding unambiguous motion
61Other Prior
- Prior on direction of lightning
62Generalization
- All computation are subject to uncertainty
(ill-posed) - This includes syntax processing, language
acquisition etc. - Solution compute with probability distributions
63Binary Decision Making
Shadlen et al.
64Race Model
- Standard theory some signal is accumulated (or
integrated) to a bound. Also known as race
models. - The signal to be integrated could be the response
of sensory neurons.
65Bayesian Strategy
- The diffusion to bound model of Shadlen et al.
66A Neural Integrator for Decisions?
- MT Sensory EvidenceMotion energy
- step
- LIP Decision FormationAccumulation of evidence
- ramp
67Diffusion to bound model
68Diffusion to bound model
- Proposed by Wald, 1947 and Turing (WW II,
classified) - Stone, 1960 then Laming, Link, Ratcliff, Smith,
. . .
69Diffusion to bound model
Criterion to answer Right
Accumulated evidencefor Rightwardandagainst
Leftward
Momentary evidencee.g.,?Spike rateMTRight
MTLeft
Criterion to answer Left
Seems arbitrary but why not?
Shadlen Gold (2004) Palmer et al (2005)
70MT responses
60 40 20 0
Height scales with coherence
Firing rate
Right
Left
Direction (deg)
71Diffusion to bound model
- Performance reaction time trade-off
72Best fitting chronometric functionDiffusion to
bound
73Predicted psychometric function Diffusion to
bound
74Average LIP activity in RT motion task
choose Tin
choose Tout
Note the clear asymmetry
Roitman Shadlen, 2002 J. Neurosci.
75Bayesian Strategy
- The Bayesian strategy in this case consists in
computing the posterior distribution given all
activity patterns from MT up to the current time,
76Bayesian Strategy
- Race models and Bayesian approach
Temporal sum
Unless is related to
77Bayesian Strategy
- Are neurons computing log likelihood?
- The difference of activity between two neurons
with preferred directions 180 deg away is
proportional to a log likelihood ratio.
78Bayesian Strategy
79Bayesian Strategy
- Is the log likelihood ration proportional to
?
Coherence level
80Bayesian Strategy
- Note that if you know , you
still dont know the log likelihood ration unless
youre given the coherence level. - Therefore, the animal cant know its confidence
level (the log likelihood ratio) unless it
estimates C - Another important point if we stop the race at a
fixed level of we stop at
different levels of log likelihood ratio
depending on the coherence. This is why
performance gets better when coherence increases,
even though we always stop at the same activity
threshold.
81Decision Making
- Does that mean the animal does not know how much
to trust its own decision? - Does that mean the brain does not encode
uncertainty or probability distribution? - Seems unlikely
82