Flash Lag

1
 How the visual system detects motion Computationa
l Neurobiology course David Eagleman Department
of Neurobiology Anatomy UT Houston Medical
School
2

• Why does the visual system need to measure motion?

• What information is available to the system?

• How the brain does it

3
Functions served by motion

• Detection
• Segmentation
• Direction of motion
• Guiding action
• Calculating time to collision

4
Structure from motion
5
Structure from motion
6
Structure from motion
7
Structure from motion
8
What information is available to the system?
Optic flow
9
What information is available to the system?
Real world motion occurs in three dimensional
(3D) space but all 3D velocities project to two
dimensional (2D) velocities across the retinas.
There is a many-to-one mapping of real world
velocities to image velocities. So a first step
in motion processing is recovering the 2D image
velocities
10
Models of direction-selectivity
Barlow and Levick model of direction selective
cells in rabbit retina
1
cell response
time
?Veto-mechanism or null-direction inhibition as
the basis for direction selectivity
?Cellular mechanism? Shunting inhibition
11
Left directionally selective cell
?Retinal cells
?Motion sensitive cell
12
Models of direction-selectivity doing it in a
dendrite
A stimulus in the preferred direction activates
retinotopic inputs from the LGN, which stimulate
the dendrites of a Meynert cell. Because inputs
farthest away take longer to activate the cell,
while those closer activate the cell more
quickly, the moving stimulus produces a
synchronized wave of excitation. Inputs near the
cell body stimulate inhibitory interneurons,
which turn off firing. Stimuli moving in the
opposite, or nonpreferred, direction encounter
inhibition first, which prevents the cell from
producing a synchronized wave of firing.
13
Models of direction-selectivity
Reichardt motion filter core computation
delay-and-compare mechanism
A
B
Note that unlike a speedometer, response will not
increase continuously with increasing velocity
instead, going beyond an optimum velocity will
decrease the response
X
Dt
14
Models of direction-selectivity
Reichardt motion filter core computation
delay-and-compare mechanism
A
B
Dt
Dt
X
X
15
Motion energy models
16
Motion energy models
Motion energy models were originally proposed
based on visual psychophysics (from many
species) (Adelson and Bergen, 1985 Watson and
Ahumada, 1985 van Santen and Sperling, 1985)
Later found to describe the behavior of
directionally selective cells in the primary
visual cortex quite well. So how does the brain
do it?
17
Spatiotemporal receptive fields
The Adelson-Bergen Motion Energy Model (1985)
x
t
18
Apparent motion
19
Apparent motion
20
Apparent motion
21
Apparent motion
22
Apparent motion
23
Spatiotemporal receptive fields
x
t
Apparent motion also stimulates spatiotemporal
receptive fields
24
The aperture problem
The same movement will be seen in the aperture
from many different object velocities
A single normal velocity is consistent with
infinitely many object velocities
25
Direction-selective cells
A cell is defined as direction selective if it
fires vigorously when the animal is presented
with a stimulus moving in one direction, and not
in the other direction. Tuning curve
(preference)
26
Different sorts of information are processed in
different neural streams
27
Area MT

• MT neurons have properties more like global
  motion perception
• MT neurons have large receptive fields (good for
  motion contrast)
• Sensitive to transparent motion (shadow moving on
  ground)
• Can use other cues (such as color) to home in on
  motion of a target
• Appear to be necessary for motion perception
  (lesion studies)
• Stimulation of MSTd neurons changes perceptual
  report

28
Vector averaging
Motion plaids are created by adding together 2
drifting gratings
Finding V1 cells respond to component motion, MT
cells respond to plaid direction
29
Higher levels of motion processing
30
Are motion-sensitive cells related to perception?

• Britten et al, 1992 The psychophysics match the
  physiology the cells perform about as well as
  the monkey
• Psychophysical and neurophysiological
• Trained monkeys to indicate direction of moving
  dots
• Recorded from single neurons in area MT
• Conclusion psychophysical judgments could be
  based on activity of small number of neurons
• Celebrini Newsome, 1995 Appropriate electrical
  stimulation biases the monkeys choices toward
  the direction of motion encoded by the neurons.
• Train monkeys to perform direction discrimination
  task. Microstimulate small groups of neurones in
  MT or MST
• Conclusion activity in areas MT and MST
  influences perception
• Motion Agnosia (motion blindness)

31
The motion aftereffect
32
The motion aftereffect
33
The motion aftereffect
Sunderland's ratio model ? ratio of activity in
two detectors tuned to opposite directions of
motion. Modified by Mather (1980). Two
superimposed random dot patterns that moved in
different directions ? resulting after effect is
in a single direction opposite to the vector sum
of the adapt directions. So direction is encoded
with a distribution of activity across a family
of orientation tuned detectors and the activities
of these detectors is combined to give a single
direction of motion.
34
Motion as a primary quality
Exner (1875) Observers could see movement
between two sparks when they were placed so close
together that the two sources could not be
resolved when triggered at the same time and when
the temporal interval was so short that pairs of
sparks at the same source could not be
distinguished from a single spark. So motion
does not require the prior computation of spatial
displacement or temporal intervals. ?Motion is
a primary sensory quality not a secondary effect.
35
Integrating information about motion takes time
The flash-lag effect A moving object, aligned
with a flashed object, appears to be displaced
proposed explanation Motion extrapolation The
visual system predicts where the moving object
will be
36
The flash-lag illusion
37
(No Transcript)
38
(No Transcript)
39
0o
40
(No Transcript)
41
Contrary to the extrapolation model, the same
initial trajectory can lead to different
flash-lag percepts
-
15
-
1
0
-
5
Perceived Displacement (deg)
0
5
1
0
1
5
o
-7
o
0
continuous
stopped
reversed
o
7
The perception attributed to an event at time
to depends on what happens in the near future
42
Prediction versus postdiction
motion extrapolation
time
motion interpolation
43
Prediction vs Postdiction
time
44
Visual perception prediction or postdiction?
45
-
15
A
-
1
0
-
5
Perceived Displacement (deg)
0
5
1
0
1
5
stopped
reversed
continuous
-
15
No initial trajectory
B
-
1
0
-
5
Perceived Displacement (deg)
0
5
1
0
Figure 1 Eagleman Sejnowski Science, 2000
1
5
46
SUMMARY

• The detection of motion is behaviorally important
• Motion seems to be a basic feature in neural
  representation (cells tuned specifically for it)
• Motion detecting neurons are necessary for proper
  motion perception
• There are several ways to build
  direction-selectivity out of biological parts
  (all involve spatial asymmetry)
• In higher cortical areas, single neurons display
  properties more like global motion perception

47
Bayesian models
P(x,y) P(x)P(yx)
P(boy,tall) P(boy)P(tallboy)
P(boy,tall) P(tall)P(boytall)
P(tall)P(boytall) P(boy)P(tallboy)
P(boytall) P(boy)P(tallboy)
--------------------- P(tall)
48
Bayesian models of motion
P(boytall) P(boy)P(tallboy)
--------------------- P(tall)
Probability
boy girl
Maximum likelihood ? pick the highest probability
49
Bayesian models of motion
P(directionretinal info) P(direction)P(retinal
infodirection)
----------------------------------------
P(retinal info)
Probability
direction
Maximum likelihood ? pick the highest probability