Computational Intelligence - May 2015 - 50

Table 1 Average difference and standard deviations between the rat's head orientations estimated from the rat-mounted
camera and from the top-mounted camera, and the accuracy of the estimated rat's head motion direction.
RouTe #1

RouTe #2

RouTe #3

RouTe #4

RouTe #5

RouTe #6

RouTe #7

RouTe #8

8.92/6.60

13.10/6.09

12.68/6.90

7.71/5.11

10.21/6.69

8.41/6.34

9.10/6.17

9.74/6.67

Trial #1

aD/SD*(Degree)
accuracy (%)

88.37

84.67

82.49

93.33

91.21

88.89

89.30

92.31

Trial #2

aD/SD (Degree)

8.68/6.55

6.86/5.97

8.44/6.81

7.74/5.33

8.19/5.52

8.58/5.95

11.54/6.18

11.38/7.14

accuracy (%)

91.03

89.77

90.39

91.65

93.59

84.17

85.21

90.67

Trial #3

aD/SD (Degree)

10.55/7.27

9.67/6.87

11.04/7.27

10.07/7.02

8.44/6.81

11.93/6.43

10.87/6.59

11.54/6.99

accuracy (%)

94.19

89.84

89.96

91.58

93.26

85.16

87.24

92.13

Trial #4

aD/SD (Degree)

8.27/7.18

13.57/7.21

8.58/5.90

8.48/6.66

11.02/6.66

8.41/5.90

9.94/6.61

8.13/6.36

accuracy (%)

86.09

85.48

88.16

94.17

91.71

89.58

91.67

89.38

* aD DenoTeS The average Difference, SD DenoTeS The STanDarD DeviaTion.

FiguRe 9 automatic cue-guided navigation.
The rat cyborg was expected to follow the
signs to reach the target face.

Table 2 Average confusion matrix for
the off-line instruction classifications
obtained using our human-like
instruction model.
%

FoRwaRd

leFT

RighT

forwarD

93.83

2.95

3.22

lefT

11.81

88.19

righT

12.37

87.63

model continuously estimates the rat
state, one estimation error may be followed by several correct estimations.
The error action caused by an error
state can be rectified by the subsequent
estimations in the model. Thus, the performance can satisfy the requirement of
the closed-loop stimulation model.
D. Evaluation of the Closed-Loop
Stimulation Model

During the evaluation of the closedloop stimulation model, we manually
controlled the rat cyborg to navigate the
four-armed maze and a small urban
planning model (Fig. 9) to collect data
for training the human-like instruction

model. After training, we employed the
unused data to perform an off-line test
of the human-like instruction model.
Next, we tested whether the closedloop stimulation model could automatically direct the rat cyborg to perform a
single action or two successive actions
given one or two object(s) and the rat's
states. Several simple scenes were designed for this experiment, where each
contained one or two object(s) and the
rat cyborg was required to perform a
single action or two actions.
To train the human-like instruction
model and to verify its performance in
an off-line test, we performed 60 trials
in the small urban planning model and
the four-armed maze. In these trials, we
placed the pictures of colored arrows
and human faces in different positions.
The rat cyborg was manually controlled
to walk toward the arrow pictures, to
turn in the directions indicated by the
arrows, and it finally reached the human
face target. During navigation, we collected the instructions issued by humans
and the videos from the rat-mounted
camera. The objects in the videos were
detected using the object detection
methods and the rat's states were
extracted. For each manually issued
instruction, we determined the synchronous rat state changes and the objects
detected to form a dataset. There were
1171 instructions in 60 trials, thus the
dataset contained 1171 samples. We
selected 574 random samples as the
training data to learn the human-like
instruction model and used the remaining 597 samples as testing data. The offline test results are shown in Table 2.

IEEE ComputatIonal IntEllIgEnCE magazInE | maY 2015

There are three types of instruction:
"Forward," "Left," and "Right." The
average confusion matrix shows that the
accuracies of classification for the three
instructions are 93.83%, 88.19%, and
87.63%, respectively.
To verify whether the closed-loop
stimulation model could automatically
direct the rat cyborg to perform a specified motion, we designed four simple
routes: a single left/right turn in the
four-armed maze and a left/right turn
followed by a right/left turn in the
small urban planning model. The pictures of arrows were placed at junctions
to indicate the motion directions. We
used two rat cyborgs and tested each rat
cyborg on the first two routes 50 times
and on the other two routes 20 times. If
the rat cyborg completed the motion
successfully, we treated it as a successful
trial. We recorded the time costs and
speed during each trial to evaluate the
efficiency of the closed-loop model.
Moreover, the same experiments were
conducted using manual control for the
purposes of comparison.
Table 3 compares the success rates
for achieving specified motions using
our automatic control method and
manual control based on four simple
tasks. In all cases, the success rates with
the closed-loop stimulation model are
very similar to those with manual control. The speed of motion completion is
also similar to the two methods. In some
cases, our method is even faster than
manual control. This indicates that the
closed-loop stimulation model can automatically control the rat cyborg to perform specified motions.

Table of Contents for the Digital Edition of Computational Intelligence - May 2015