Information Technology Reference
In-Depth Information
Probe RF
Item Summary
Probe Stims
itm
err
%tot
%itm
correl
uniq
2
3
0
19
0.64
1.85
97.76
4.22
1
65
2.22
6.34
91.85
7.74
2
3
1
0
0
1
2
71
2.42
6.93
92.70
8.37
3
35
1.19
3.41
96.44
5.52
4
42
1.43
4.10
95.66
5.27
5
42
1.43
4.10
95.78
5.53
6
23
0.78
2.24
97.80
4.65
LGN
7
42
1.43
4.10
95.96
4.63
8
70
2.39
6.83
92.90
8.70
Figure 8.19: Organization of the probe receptive fields for
all 4 probe stimuli as displayed in one plot. The 4 probes (0-
3) were each swept across the lower-left quadrant of the LGN
input, and the activations of V2 and V4 units recorded. For
each V2/V4 unit, its response when the probe was in a given
location (indexed by the lower left-hand corner of the probe)
is plotted in the probe RF, with the quadrants of the probe RF
arranged in the same way the probes themselves are.
9
70
2.39
6.83
93.38
6.08
10
35
1.19
3.41
96.47
5.56
11
43
1.46
4.19
96.09
4.96
12
61
2.08
5.95
94.25
7.97
13
28
0.95
2.73
96.98
4.59
14
41
1.40
4.00
95.37
5.31
15
62
2.11
6.05
94.06
6.21
16
57
1.94
5.56
94.49
7.47
17
72
2.46
7.03
92.03
6.37
18
1024
34.99
100
73.29
9.43
You should observe much greater levels of spatial in-
variance in the V4 units (figure 8.20) compared to the
V2 units, as we would expect from previous analyses.
Several of these V4 units responded to a single feature
across most of the locations it was presented, which cor-
responds to a single yellow quadrant of nearly-solid ac-
tivity in the probe display. Other units responded to
two different features in this invariant fashion, corre-
sponding to two quadrants of activity. Some of the units
even appear to respond to all probes across all locations.
Given that the probes all involved a simple combina-
tion of a vertical and horizontal line, these units that did
not discriminate the different probes are likely encoding
less specific aspects about the vertical-horizontal line
junction (and even just the very fact of such a junction
in the least specific case).
19
1024
34.99
100
55.47
8.19
size
Size Summary
0
1012
34.58
19.76
88.77
6.33
1
614
20.98
11.99
90.03
6.34
2
612
20.91
11.95
95.34
6.34
3
688
23.51
13.43
93.62
6.35
Tab le 8 . 1 : Summary of sweep testing results prior to gener-
alization training on novel items (18 and 19). itm gives the
object by number, err is the total number of errors across all
positions and sizes for that object, %tot is the percentage of er-
rors for that item out of total errors, %itm is the percentage of
errors out of total presentations of the item, correl is the aver-
age correlation between subsequent V4 representations for the
item, and uniq is the average number of unique V4 represen-
tations across all 256 different locations. The last four rows
show the same kinds of statistics tabulated instead for the four
different sizes — more errors are made on the smallest size.
The uniq value for the novel items (18 and 19) indicate rel-
atively invariant representations without any specific training
on these items.
Sweep Analysis
A similar way of testing the representations in the net-
work is to sweep all of the actual object stimuli system-
atically across all of the positions in the input, at all the
different sizes, and record some statistics on the result-
Search WWH ::




Custom Search