of 9
Supporting Information
Navalpakkam et al. 10.1073/pnas.0911972107
Decision Models
Here, we explain the derivation of the global likelihood terms in
Eqs.
3
5
. Let
P
ð
T
x
¼
H
j
a
!
Þ
denote the posterior probability of
target H
s presence at location
x
given the vector of sensory ob-
servations at all locations in the display,
a
!
¼f
a
1
...
a
n
g
. Each
display consists of target H, V, and
n
2 distractors. Thus, the
posterior probability of H
s presence at location
x
can be expressed
as the joint probability of H
s presence at
x
, target V
s presence at
any other location
y
x
, and distractor D
s presence at the re-
maining locations
z
x
,
y
, marginalized over all choices of
y
. The
posterior probabilities can then be expressed as a product of the
likelihood and prior terms. We assume that target V can occur at
any location
y
x
with equal probability,
P
ð
T
y
¼
V
j
T
x
¼
H
Þ¼
1
n
1
.
Thus, we get the following:
P
ð
T
x
¼
H
j
a
!
Þ¼
y
x
P
ð
T
x
¼
H
;
T
y
¼
V
;
T
z
x
;
y
¼
D
j
a
!
Þ
[S1]
P
ð
a
!
j
T
x
¼
H
Þ
P
ð
T
x
¼
H
Þ¼
y
x
P
ð
a
!
j
T
x
¼
H
;
T
y
¼
V
;
T
z
x
;
y
¼
D
Þ
×
P
ð
T
y
¼
V
j
T
x
¼
H
Þ
P
ð
T
x
¼
H
Þ
[S2]
P
ð
a
!
j
T
x
¼
H
Þ¼
1
n
1
y
x
P
ð
a
!
j
T
x
¼
H
;
T
y
¼
V
;
T
z
x
;
y
¼
D
Þ
[S3]
In
Eq.
S3
,
P
ð
a
!
j
T
x
¼
H
Þ
denotes the global likelihood of target
H
s presence at location
x
, based on the sensory observations at
all locations in the display,
a
!
¼f
a
1
...
a
n
g
. Let
P
(
a
x
|
T
x
=
H
)
denote the local likelihood of H
s presence at location
x
based on
the sensory observation at that single location,
a
x
.
The global likelihood term on the right-hand side in
Eq.
S3
can
be expressed as a product of local likelihoods, as shown in
Eq.
S4
. This completes the derivation of Eq.
3
. Eqs.
4
and
5
are
derived similarly.
P
ð
a
!
¼f
a
1
...
a
n

T
x
¼
H
Þ¼
1
n
1
y
x
P
ð
a
x
j
T
x
¼
H
Þ
×
P
ð
a
y
j
T
y
¼
V
Þ
Π
z
x
;
y
P
ð
a
z
j
T
z
¼
D
Þ
[S4]
Experiment 1: Orientation
IndividualSubjects
DataComparedwithModelPredictions.
Fig. 2
A
D
shows data from subject S1. Here, we show the data from the
remaining subjects S2 through S6. As seen in
Figs. S1
S5
, be-
havior of all subjects is consistent with the predictions of M3.
Thus, subjects in experiment 1 behave as reward maximizers that
saccade to the location of the maximum expected reward.
In experiment 1 (and others),
fi
xations could land on either
target H, target V, or distractor D.
Fig. S9
shows the distribution
of
fi
xations of subject 1 on target H, target V, and distractor D in
different value and feature-contrast conditions. Note that the
average probability of
fi
xating a distractor is 8.1% (SD = 5.6%)
and that it is not affected by the relative value or feature-contrast
of the targets [no effect of value using paired
t
tests at a sig-
ni
fi
cance level of 0.05; no effect of feature-contrast using paired
t
tests at a signi
fi
cance level of 0.05; no correlations between %
fi
xations on the distractor and target
s relative value (
R
2
= 0.01)
or feature-contrast (
R
2
= 0.04)]. For this reason, all subsequent
analyses focus on
fi
xations to one of the targets (e.g., H).
Fig. S9
also shows that in many conditions, subjects chose
fl
exibly be-
tween targets H and V (e.g.,
Fig. S9
,
Upper Left
, when target H is
twice as salient but V is twice as valuable;
Upper Right
, when both
targets are equally salient and valuable;
Lower Left
, when target
H is twice as valuable but V is twice as salient). This shows that
subjects used a dynamic strategy of choosing
fl
exibly different
targets between trials rather than a static strategy that they may
have acquired through ample practice or training in a condition.
Saccade Latency.
How do value and feature-contrast affect latency
to the
fi
rst saccade?
Fig. S8
shows the data pooled over all six
subjects. The data show two main trends: (
i
) saccade latency de-
creases as the target
s feature-contrast increases (two-way
ANOVA shows a signi
fi
cant effect of feature-contrast:
F
(6, 18) =
898,
P
<
0.01) until it reaches a mean latency of
330 ms (the
highest latency of 500 ms occurs when the target
s feature-contrast
is very low and indicates that it is never
fi
xated within 500 ms), and
(
ii
) saccade latency decreases as the target
s relative value in-
creases (two-way ANOVA shows a signi
fi
cant effect of value:
F
(3,
18) = 28,
P
<
0.01). Thus, value and feature-contrast affect the
initial gaze and its latency. The shortest latency of around 330 ms
followed by a
fi
xation duration of at least 100 ms (required to earn
the reward) indicates that subjects mostly made one saccade
during the 500 ms of presentation in experiment 1.
Individual Subjects
Data for Experiment 2: Intensity
The reward maximization behavior observed in experiment 1 on
oriented stimuli extends to experiment 2 on luminance stimuli.
Fig. 3 and
Figs. S6
and
S7
show the data from three individual
subjects.
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
1of9
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
% fixat i ons
H
% fixat i ons
H
v
H
/v
V
= 0.5
χ
χ
χ
χ
χ
χ
χ
χχχ
χ
χ
2
(M1)=7.63,
2
(M2)=5483.16,
2
(M3)=1.01
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 1
2
(M1)=6.88,
2
(M2)=46.88,
2
(M3)=2.14
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
% fixations H
% fixations H
v
H
/v
V
= 2
2
(M1)=35.41,
2
(M2)=11.61,
2
(M3)=0.71
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 4
2
(M1)=72.75,
2
(M2)=13.54,
2
(M3)=1.84
A
0
20
40
60
80
100
0
20
40
60
80
100
M3:
2
=0.96, slope=0.96
M2:
2
=0.51, slope=0.57
M1:
2
=0.81, slope=0.94
subject’s detection rates
Predicted detection rates
B
R
R
R
Fig. S1.
Experiment 1, subject S2 (author). (
A
) Subject
s data (black dots with error bars denoting SEM) and the psychometric functions predicted by the three
models (M1: light gray, M2: dark gray, M3: red). Each panel denotes a different ratio of values of targets H vs. V. The
χ
2
goodness-of-
fi
t statistic (comparing
how well each model
fi
ts the data) is shown in the title. Across all conditions tested, M3 (the ideal observer)
fi
ts the subject
s data better than M1 and M2. (
B
)
Predictions of model M3 correlate well with the subject
s data (pooled across all value conditions).
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
2of9
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
% fixations H
% fixations H
% fixations H
% fixations H
v
H
/v
V
= 0.5
χ
χ
χ
χ
χ
χ
χ
χ
χ
χ
χ
χ
2
(M1)=7.08,
2
(M2)=5389.96,
2
(M3)=1.43
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 1
2
(M1)=2.45,
2
(M2)=27.92,
2
(M3)=1.15
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 2
2
(M1)=49.01,
2
(M2)=7.86,
2
(M3)=3.48
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 4
2
(M1)=65.21,
2
(M2)=2.14,
2
(M3)=1.37
A
020406080100
0
20
40
60
80
100
M3:
2
=0.97, slope=1.01
M2:
2
=0.50, slope=0.61
M1:
2
=0.79, slope=0.96
subject’s detection rates
Predicted detection rates
B
R
R
R
Fig. S2.
Experiment 1, subject S3; similar to
Fig. S1
.
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
3of9
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
% fixations H
% fixations H
% fixations H
% fixations H
v
H
/v
V
= 0.5
2
(M1)=9.25,
2
(M2)=3934.10,
2
(M3)=2.17
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 1
2
(M1)=2.62,
2
(M2)=22.32,
2
(M3)=2.97
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 2
2
(M1)=31.13,
2
(M2)=9.93,
2
(M3)=1.43
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 4
2
(M1)=54.49,
2
(M2)=6.00,
2
(M3)=1.76
0
20
40
60
80
100
0
20
40
60
80
100
M3:
2
=0.94, slope=1.00
M2:
2
=0.55, slope=0.64
M1:
2
=0.76, slope=0.95
subject’s detection rates
Predicted detection rates
A
B
χ
χ
χχ
χ
χ
χ
χ
χ
χ
χ
χ
R
R
R
Fig. S3.
Experiment 1, subject S4; similar to
Fig. S1
.
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
4of9
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
feature−contrast H/V
% fixations H
% fixations H
% fixations H
% fixations H
v
H
/v
V
= 0.5
2
(M1)=11.91,
2
(M2)=3341.05,
2
(M3)=2.14
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
feature−contrast H/V
v
H
/v
V
= 1
2
(M1)=2.70,
2
(M2)=21.51,
2
(M3)=1.79
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
v
H
/v
V
= 2
2
(M1)=38.68,
2
(M2)=5.07,
2
(M3)=10.79
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
v
H
/v
V
= 4
2
(M1)=109.08,
2
(M2)=5.84,
2
(M3)=3.58
0
20
40
60
80
100
0
20
40
60
80
100
M3:
2
=0.91, slope=1.02
M2:
2
=0.59, slope=0.70
M1:
2
=0.73, slope=0.97
subject’s detection rates
Predicted detection rates
A
B
χ
χ
χ
χ
χχ
χ
χχ
χ
χ
χ
R
R
R
Fig. S4.
Experiment 1, subject S5; similar to
Fig. S1
.
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
5of9
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
% fixations H
% fixations H
% fixations H
% fixations H
v
H
/v
V
= 0.5
2
(M1)=10.49,
2
(M2)=1835.80,
2
(M3)=2.83
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 1
2
(M1)=2.98,
2
(M2)=30.72,
2
(M3)=4.26
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 2
2
(M1)=28.26,
2
(M2)=9.24,
2
(M3)=0.93
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 4
2
(M1)=74.47,
2
(M2)=7.00,
2
(M3)=2.12
020406080100
0
20
40
60
80
100
M3:
2
=0.92, slope=0.97
M2:
2
=0.55, slope=0.60
M1:
2
=0.74, slope=0.89
subject’s detection rates
Predicted detection rates
A
B
χ
χ
χχχχ
χ
χχ
χ
χ
χ
R
R
R
Fig. S5.
Experiment 1, subject S6; similar to
Fig. S1
.
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
6of9
0.2 0.33 0.50.71 1 1.4 2
5
0
20
40
60
80
100
feature−contrast L/H
% fixations L
% fixations L
% fixations L
v
L
/v
H
= 1
2
(M1)=1.40,
2
(M2)=63.47,
2
(M3)=7.63
0.2 0.33 0.50.71 1 1.4 2
5
0
20
40
60
80
100
feature−contrast L/H
feature−contrast L/H
v
L
/v
H
= 2
2
(M1)=152.72,
2
(M2)=17.66,
2
(M3)=2.44
0.2 0.33 0.50.71 1 1.4 2
5
0
20
40
60
80
100
v
L
/v
H
= 4
2
(M1)=388.99,
2
(M2)=11.60,
2
(M3)=2.42
020406080100
0
20
40
60
80
100
subject’s detection rates
Predicted detection rates
M1:
2
=0.68, slope=0.90
M2:
2
=0
.
68, slope=0.53
M3:
2
=0.95, sl o pe=0.89
A
B
χ
χ
χ
χ
χ
χ
χ
χ
χ
R
R
R
Fig. S6.
Experiment 2, subject S2; similar to
Fig. S1
but using brightness intensity as the feature.
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
7of9
0.2
0.50.71 1 1.4 2
5
0
20
40
60
80
100
feature−contrast L/H
% fixations L
% fixations L
% fixations L
v
L
/v
H
= 1
2
(M1)=6.73,
2
(M2)=45.25,
2
(M3)=1.73
0.2
0.50.71 1 1.4 2
5
0
20
40
60
80
100
feature−contrast L/H
v
L
/v
H
= 2
2
(M1)=186.31,
2
(M2)=13.84,
2
(M3)=0.68
0.2
0.50.71 1 1.4 2
5
0
20
40
60
80
100
feature−contrast L/H
v
L
/v
H
= 4
2
(M1)=325.44,
2
(M2)=10.78,
2
(M3)=5.86
0 20406080100
0
20
40
60
80
100
subject’s detection rates
Predicted detection rates
M1:
2
=0.53, slope=0.87
M2:
2
=0.69, slope=0.57
M3:
2
=0.96, slope=0.94
A
B
χ
χχ
χ
χ
χ
χ
χ
χ
R
R
R
Fig. S7.
Experiment 2, subject S3; similar to
Fig. S1
but using brightness intensity as the feature.
0.059
0.2
0.5
1
2
5
17
320
340
360
380
400
420
440
460
480
500
feature−contrast H/V
RT ( m s)
0.5
1
2
4
v
H
/ v
V
Fig. S8.
Saccadic latency for experiment 1. This
fi
gure shows the time to saccade to the horizontal target as a function of its relative value and feature-
contrast. Saccades become faster as the target
s value and feature-contrast increase. RT, reaction time.
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
8of9
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
%fixations
%fixations
%fixations
%fixations
v
H
/v
V
= 0.5
H
V
D
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 1
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 2
0.059
0.2
0.5
1
2
5
17
0
20
40
60
80
100
feature−contrast H/V
v
H
/v
V
= 4
Fig. S9.
Experiment 1, subject S1. Fixations to the distractor (blue), target H (black), and target V (green) for different value and feature-contrast condit
ions in
experiment 1. Although the
fi
xations to the targets vary systematically with value and feature-contrast, the
fi
xations to the distractor appear to be unaffected
(details provided in
SI Text
).
Navalpakkam et al.
www.pnas.org/cgi/content/short/0911972107
9of9