A Synthetic Nervous System Model of the Insect Optomotor Response
315
The SNS model of the optic lobe runs on a laptop computer separate from the
Raspberry Pi. The Pi processes and sends the data to the laptop over a serial connection
by the following process. 64 × 64-pixel grayscale video is recorded and dissected for
static images at about 25 Hz. Such low image resolution reduces serial traffic and is
consistent with insects’ comparatively low visual resolution [1]. This image is sorted
into 64, 5.6 degree-wide angular “bins” along the azimuth, wherein each bin’s intensity
value is the average of all pixels it encapsulates (illustrated in Fig. 1C). Note that this
system only “sees” along one axis, the azimuth; it cannot detect changes in the elevation
direction. When a bin’s intensity value changes in subsequent images, it is flagged for
transmission over serial to the SNS in the next data sentence (Fig. 1D). Such a system
reduces the length of sentences and thus increases the system’s bandwidth.
Visual information from the camera is transduced into neural inputs at the retina
layer of the optic lobe model (Fig. 1E). Each angular bin in the camera’s field of view
has a corresponding retinal cell. The average grayscale intensity of each bin is mapped
to an applied current for the retinal cell with black mapping to a current of 0 nA, white
mapping to a current of 20 nA, and intermediate values mapping in a graded way. Each
retinal cell represents the first layer of a columnar network that enhances contrast, detects
changes in pixels, and sums these changes over the field of view to compute wide-field
visual velocity. This network is described in detail in Sect. 2.3.
The connection between the robot and the SNS is bidirectional. The SNS possesses
motor neurons (MNs) for the “neck” of the robot to rotate the camera around the vertical
axis. There are two MNs, each of which rotates the neck in the opposite direction. In our
previous work, we found that the motor output of small animals can be approximated
by using the sum of the MN voltages to set the servo’s speed and using the difference
of the MN voltages to set the servo’s equilibrium angle [10]. If the commanded angle or
speed of the servo changes, then a new command sentence is sent over serial from the
SNS controller to the robot.
2.2 Synthetic Nervous System Organization and Design
The SNS model of the optic lobe is implemented with Animatlab [11] and its Robotics
Toolkit [12]. Animatlab is an open-source 3D neuromechanical simulation software tool
for simulating biologically-inspired organisms, robots, and neural networks. We wished
to build a network that is biologically plausible but possible to run in real time. While the
optic lobes of different insect species have behavior-specific visual processing networks,
the overall structural organization is common for several kinds of arthropods including
moths, flies, crabs, and mantises [13]. Since we are modeling a fundamental behavior
observed in many species, we have combined information from multiple species. However, modeling more species-specific behaviors (e.g. mantis prey capture) would require
modifying the network in species-specific ways.
Neural Modeling Techniques. Our network model is composed of integrate-and-fire
neurons [14]. To simplify the description of the neurons, we write the equations in terms
of U , the membrane voltage above the neuron’s rest potential [15]. In addition, each
neuron has a membrane conductance G m = 1 and a constant spiking threshold θ = 1.
315
The SNS model of the optic lobe runs on a laptop computer separate from the
Raspberry Pi. The Pi processes and sends the data to the laptop over a serial connection
by the following process. 64 × 64-pixel grayscale video is recorded and dissected for
static images at about 25 Hz. Such low image resolution reduces serial traffic and is
consistent with insects’ comparatively low visual resolution [1]. This image is sorted
into 64, 5.6 degree-wide angular “bins” along the azimuth, wherein each bin’s intensity
value is the average of all pixels it encapsulates (illustrated in Fig. 1C). Note that this
system only “sees” along one axis, the azimuth; it cannot detect changes in the elevation
direction. When a bin’s intensity value changes in subsequent images, it is flagged for
transmission over serial to the SNS in the next data sentence (Fig. 1D). Such a system
reduces the length of sentences and thus increases the system’s bandwidth.
Visual information from the camera is transduced into neural inputs at the retina
layer of the optic lobe model (Fig. 1E). Each angular bin in the camera’s field of view
has a corresponding retinal cell. The average grayscale intensity of each bin is mapped
to an applied current for the retinal cell with black mapping to a current of 0 nA, white
mapping to a current of 20 nA, and intermediate values mapping in a graded way. Each
retinal cell represents the first layer of a columnar network that enhances contrast, detects
changes in pixels, and sums these changes over the field of view to compute wide-field
visual velocity. This network is described in detail in Sect. 2.3.
The connection between the robot and the SNS is bidirectional. The SNS possesses
motor neurons (MNs) for the “neck” of the robot to rotate the camera around the vertical
axis. There are two MNs, each of which rotates the neck in the opposite direction. In our
previous work, we found that the motor output of small animals can be approximated
by using the sum of the MN voltages to set the servo’s speed and using the difference
of the MN voltages to set the servo’s equilibrium angle [10]. If the commanded angle or
speed of the servo changes, then a new command sentence is sent over serial from the
SNS controller to the robot.
2.2 Synthetic Nervous System Organization and Design
The SNS model of the optic lobe is implemented with Animatlab [11] and its Robotics
Toolkit [12]. Animatlab is an open-source 3D neuromechanical simulation software tool
for simulating biologically-inspired organisms, robots, and neural networks. We wished
to build a network that is biologically plausible but possible to run in real time. While the
optic lobes of different insect species have behavior-specific visual processing networks,
the overall structural organization is common for several kinds of arthropods including
moths, flies, crabs, and mantises [13]. Since we are modeling a fundamental behavior
observed in many species, we have combined information from multiple species. However, modeling more species-specific behaviors (e.g. mantis prey capture) would require
modifying the network in species-specific ways.
Neural Modeling Techniques. Our network model is composed of integrate-and-fire
neurons [14]. To simplify the description of the neurons, we write the equations in terms
of U , the membrane voltage above the neuron’s rest potential [15]. In addition, each
neuron has a membrane conductance G m = 1 and a constant spiking threshold θ = 1.