Biomedical Engineering Reference
In-Depth Information
single-trial images associated with a given color into a single class, regardless of
stimulus orientation. We used data from 31 of the 32 imaging blocks to train
an SVM classifier and test the SVM with data from the remaining block. Each
imaging block consisted of one presentation of each stimulus. In the training phase,
a linear SVM [29] was derived from the collection of training pairs of single-trial
image and its associated color. In the testing phase, the SVM classifier inferred
the stimulus color that was associated with each test single-trial image, and the
inferred color was compared with the actual one. This training-testing procedure
was repeated 32 times, with each repetition using a different imaging block for
testing. Across the total of 512 tests, the SVM gave 430 (84%) correct inferences
about the stimulus color. So the accuracy of decoding stimulus color was 84%,
significantly higher than what it would be by chance (1 in 4 colors, or 25%).
Similarly, the accuracy for decoding orientation was 74%. Therefore, each single-
trial image contained a significant amount of information about stimulus color and
orientation.
To map the spatial distribution of information about color and orientation
across the visual cortex, a ''search light '' approach was used [30]. In this approach,
a ''search light'' of 494
81 pixels) scanned through the entire
imaged area, shifting its center location by one pixel at each stop. At each location,
the SVM analysis described above was applied to the highlighted region, and the
resulting accuracy was assigned to the pixel at the given center location. This
approach produced two accuracy maps, one for stimulus color, and the other for
stimulus orientation [Figures 10.4(a, b), respectively].
Figure 10.4(a, b) shows that information about color and orientation had
different patterns of spatial distribution. To compare these patterns directly, we
selected those pixels that had a value of
×
494 microns (81
×
50% in accuracy map of color or ori-
entation, and marked them with different colors in Figure 10.4(c) (black for color
decoding, light gray for orientation decoding, and white for both). This figure
suggests that the information about color and orientation was concentrated in
different regions that were largely segregated from each other.
The arrows in Figure 10.4(c) point to the same locations as those in Figure
10.1(c), and thus mark the locations of two CO stripes. It is evident that the
regions containing large amount of color information were centered in the thin
stripes, whereas the region with orientation information were located between the
thin stripes.
>
10.3 Discussion
Our results suggest that information about stimulus color and orientation is coded
by the spatial pattern of responses in different CO stripes of V2. Previous imaging
studies also suggested that the thin CO stripes and the interstripes are involved
in coding color and orientation information, respectively. Cortical columns that
were selective for stimulus orientation were found in the interstripes and thick
stripes; whereas columns that responded preferentially to color stimuli were found
in the thin stripes [20, 28]. Furthermore, the thin stripes were found to contain
Search WWH ::
Custom Search