Computational Intelligence - May 2015 - 47

2) Human-like Instruction Model
In the closed-loop stimulation method,
the human-like instruction model issues
an instruction, given a rat state and an
object. This method operates in a similar
manner to humans when they encounter a similar situation during the manual
control of a rat cyborg.
In the manual control process, after
the instruction C i -1 is issued to the rat
cyborg, its posture changes from the state
S i -1 to the next state S i . Next, humans
observe the change in the rat state and
determine the current instruction C i to
adjust the incorrect action or to reward
the correct action. Thus, the state change
is an important factor when deciding the
current instruction. In addition, the current object affects the selection of the
instruction. Different objects are related
to different motion expectations for the
rat cyborg. Thus, different instructions
will be sent even if the same state change
is observed for different objects. Therefore, the current motion expectation is
treated as another factor when deciding
the current instruction.
We assume that the state change
extracted by our method is DS i =
{i i - i i -1, Vi - Vi -1}, t h e c u r r e n t
motion expectation E i indicated by the
object detection result is one of Forward
(0), Left Turn (-1), and Right Turn (1),

State 1

State 2

and the current instruction issued by
humans is C i . We use X i = (DS i, E i) as
input features and C i as output labels to
train a support vector machine (SVM)
classifier to construct the human-like
instruction model. Given a set of l training data points (X i, C i), i = 1, 2, f, l,
the SVM aims to learn a classifier for
the input data X with the form:
C = sign(w T X + b),

Target

i
Image Plane

(6)

Miniature
Camera

where w and b are parameters
obtained by solving the following optimization problem:
l

min 1 w 2 + a / p i,
2
i =1
s.t. C i (w T X i + b) $ 1 - p i and p i $ 0,
(7)
where p is a slack variable and the
parameter a balances the weights
between the two terms. Because there
are three possible values for an instruction: C (Left (-1), Right (1), or Forward (0)), we build a three-class classifier
as the human-like instruction model
using a one-against-all scheme.
The training data were collected
during the manually controlled navigation of the rat cyborg. Both the control
instructions and the videos from the ratmounted camera were recorded. The rat
motion state S i that corresponded to

State 3

State 4

D (Scene
Depth)

Figure 4 Rat head orientation estimation.

each instruction C i was extracted from
the videos and used to compute the rat
state change DS i . The action expectation E i was obtained based on the
object detection results. In the testing
stage, we computed the real-time state
change DS and motion expectation E
of the rat cyborg. The instruction for the
rat cyborg was then obtained based on
the classification results produced using
our model. Figure 5 shows the training
and testing framework employed by our
human-like instruction model.
VI. Experiments

In this section, we present details of the
rat surgery, including the electrodes and

State 5

Rat State Change & Detection Result

Object Detection
Rat State Extraction
Instruction
SVM

Instruction Data

Instruction 1:
Left

Instruction 2:
Forward

Instruction 3:
Right
Model Learning

Instruction 4:
Forward

Instruction 5:
Forward
Automatic Navigation

Figure 5 Human-like instruction model.

MAY 2015 | IEEE CoMputAtIonAl IntEllIgEnCE MAgAzInE

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