Neuromorphic VLSI Modeling of Bat Echolocation

Neuromorphic VLSI Modeling of Bat Echolocation
Timothy K. Horiuchi
“Seeing” in the Dark: Neuromorphic
VLSI Modeling of Bat Echolocation
includes echolocation. It is most comchoes are a major part of our
monly associated with underwater sensdaily auditory lives. Yet, very
ing in submarines and dolphins, but the
few of us are even aware that
most common example around us is the
we use them constantly to
behavior of the insect-eating bats (see
understand the space around
Figure 1) we find in the dark skies on a
us. The bats of the suborder Microwarm summer evening almost anychiroptera have made echo detection
where in the world. Although moderntheir primary sensory modality for
day sonar and radar are superior in
guiding agile flight in cluttered, threemany ways, the study of biosonar
dimensional (3-D) environments and
for capturing flying insects in
complete darkness.
The attempt to understand
this fascinating sensory modality
has captured the attention of
researchers in fields as diverse as
psychology, biology, physics,
electrical engineering, and aeronautics since the discovery in the
1790s that bats could fly in cluttered rooms even when blinded.
The development of underwater
sonar during World War I, as
well as that of radar during
World War II, has driven many
advances in signal processing [FIG1] The big brown bat, Eptesicus fuscus.
techniques and technology that
remains important to engineers for
continue to occupy engineers today.
many reasons. Even backed by powerful
While modern sonar and radar processmathematical tools and decades of expeing systems are constructed using large
rience, our best underwater sonar syssensor arrays that have incredible imagtems still do not rival some of the
ing and localization capabilities, the
perceptual capabilities of dolphins.
question of what bats experience and
Using an airborne echolocation system,
how they process echo data with only
bats have incredible aerial agility, flying
two ears and a pea-sized brain remains a
in complete darkness through branches
major mystery. The exploration of this
and caves while hunting evasive insects.
question and the attempt to construct a
These animals perform such tasks with
functional model of the bat’s neural sigtotal power consumption (including
nal processing using neuromorphic very
flight) measured in watts, not hundreds
large scale integration (VLSI) techof watts [1], [2]. In addition, both bats
niques and robotics have provided an
and dolphins live in very social environinteresting framework for our laboratoments, using echolocation in group sitry’s research program.
uations without any obvious problems
The acronym SONAR comes from
with interference. These capabilities are
SOund NAvigation and Ranging, which
exactly what engineers developing
micro-aerial vehicles (MAV) are trying
to achieve. Such devices need to operate
in environments where global positioning system (GPS) signals are unreliable
and obstacles are unmapped. While bat
echolocation is a relatively short-range
sensory modality due to signal attenuation, it is appropriate for the close-quarters navigational problems and flight
speeds (1 to 6 m/s [1]) that bats
(and MAVs) encounter.
In spite of the phenomenal
growth in the usefulness, connectivity, and computational
power of the conventional digital computer, we are still not
very close to building an intelligent machine. The more we
learn about the brain, the more
it is clear that the computations
being performed and the hardware being used are quite different
computers. Our best efforts at speech
recognition, vision-based navigation,
and limbed motor control, while functional, are still basic and fragile, consume large amounts of power, and
ultimately have become increasingly
based on biological concepts. Ironically,
we are literally surrounded by functional examples of the very type of intelligent machine we would like to build.
While most creatures can determine
the direction of an arriving sound, bats
extend this ability by emitting their
own sound, which produces echoes off
of objects that are then localized
phology. Using commercial
(Figure 2). The time of
silicon foundries, large
flight (or echo delay) proarrays of analog and digital
vides information about
circuits can be (relatively)
range, and changes in the
inexpensively fabricated on a
spectrum and temporal
single chip to perform the
structure can provide informassively parallel signal promation about the shape,
cessing known to occur in
size, and motion of objects.
neural structures like the
The exact selection of the
retina [6]–[8], the cochlea
sound to be emitted repre[9]–[13], the auditory midsents an active choice in
[FIG2] Bat echolocation relies on the reflection of emitted ultrasonic
brain [14], the cortex [15],
optimizing the measure- sound pulses off objects in the environment. By varying the temporal
[16], the spinal cord [17],
ment of one feature over and spectral properties of the pulses, as well as their timing,
and other sensorimotor
another. For example, big echolocating bats appear to actively analyze their world as they fly.
structures [18]–[20]. By utibrown bats emit a short,
lizing dedicated parallel anaharmonic, downward sweep
log circuits, low-precision computations
a long history in the demonstration of
in the ultrasonic frequency range and
can be performed in real time (or faster
our understanding of biological sysactively modulate the bandwidth, loudthan real time), with power consumptems. In most cases, however, these two
ness, duration, and repetition rate.
tion that is many orders of magnitude
endeavors have been pursued separately
From a neuroscience perspective,
less than in a general-purpose computdue to a lack of knowledge of the
bats are remarkable due to the very
er. While the challenge of translating
underlying neurobiology at the cellular
short time scales at which their sensory
these devices into useful neural modeland systems levels (neuroscience), limisystems must operate. The barrage of
ing tools remains, the speed of computations of electronic realizations (fabridetectable returning sonar echoes from
tation and the small physical size
cation technology), shortcomings of
a bat’s near environment lasts about 30
promise to enable real-time modeling of
available robotic technology (materials,
ms following a sonar emission, with any
complex sensorimotor interactions that
actuators, batteries), or the limited
one target’s echo lasting a few millisecwere previously impossible (e.g.,
appreciation of the critical synergy
onds. At this time scale, any one neuron
echolocation-based flight control).
between brain and body. Continuing
(the putative computing element of the
VLSI-based neural models also stand
advances in the fabrication of microbrain) has the opportunity to fire only
out as useful tools where a large range
electronics, recent advances in neuroone or two action potentials (voltage
of time constants are desired; they are
science, and the latest consumer focus
pulses, or “spikes”) to represent the
suited for use in spike-based synaptic
on low-power, portable electronics are
echo. Unlike the “traditional” view of
learning rules or in spike-based models
driving technologies that will create
neural processing in the neocortex,
of motor control, where neuron-tonew opportunities in biologically
where action potentials from a neuron
world interactivity is desired.
inspired robotics. Hardware implemenare integrated over tens to hundreds of
tations provide a reality check for commilliseconds to estimate a firing rate,
putational models that attempt to
the bat must rely on tens to hundreds of
While roboticists have been using and
explain how neural circuits can control
neurons that respond with only a few
researching sonar for decades, there
behavior in realistic sensorimotor envispikes. In such neural circuits, the
have only been a handful of efforts to
ronments. Such implementations can
details of action potential timing, neutruly mimic the behavior of the bat.
also create new technology and devices
ron interconnection (synapse) dynamFewer still have attempted to model
for commercialization.
ics, and the internal dynamics of a
the neural circuitry underlying bat
While early efforts to model neural
neuron become extremely important. In
echolocation in a closed-loop fashion.
circuits with discrete electronic compospite of all this behavioral and sensory
Two recent examples include work to
nents provided limited insight due to
specialization, the bat brain is still
understand the computational signifismall numbers of cells and functionaliorganized like other mammalian brains;
cance of scanning pinnas movements
ty, the development of a toolbox of anait provides a valuable comparative viewof the horseshoe bat [21] and the new
log VLSI primitives for describing
point of mammalian auditory processEuropean CIRCE project (www.circeneural structures launched a field of
ing in general. to build a bionic bat head
engineering known as neuromorphic
with initial results in analyzing pinnas
VLSI [3]–[5]. This approach involves the
shape [22].
design of analog and digital VLSI cirROBOTIC IMPLEMENTATION
In the Computational Sensorimotor
cuits that mirror neural algorithms in
Electronic models of neural systems
Systems Laboratory at the University of
both signal representation and morand robot models of behavior both have
Maryland, we are focused on the neural
circuitry underlying bat echolocation
to understand and distill the fundamental computations performed by the
bat auditory system. We aim to develop
low-power neuromorphic VLSI circuits
(fabricated through the MOSIS
Service) that mimic this remarkable
system in real time. We plan to demonstrate the success of our understanding
through the construction of a small,
sonar-guided aircraft. The motivating
theme of our laboratory is the investigation of how sensory information is
extracted by neural systems and how
neural activity is ultimately transformed into system-level behavior. In
the following sections, we highlight
some of the bat echolocation subsystems we are studying and describe
what challenges we face in implementing them in silicon.
[FIG3] An example binaural sonar head
with two ultrasonic microphones
mounted on top, with an ultrasonic
loudspeaker below. The assembly is
shown mounted on a rotating platform.
Neurons +
[FIG4] LSO neurons receive excitation
from one ear and inhibition from the
other ear. Differences in logarithmically
encoded input strength and thresholds
produce sensitivity to different sound
Echolocation begins with the bat (or
robot) emitting a short (e.g., 2 ms), but
loud, ultrasonic vocalization. This
sound is transmitted out from the head
and produces echoes from small objects
at distances as far away as 5 m, a significant distance given the small size of the
animal. After the echoes return to the
bat (or robot), the head, snout, and pinnas (or microphone sensitivity curves)
all combine to produce directionally
dependent spectral filtering for each of
the two ears (or microphones). While
this binaural, head-related transfer
function (HRTF) has been well characterized in humans, existing measurements for bats are coarse, and only a
few examples exist. A number of groups
have recently been pursuing these
measurements (such as the CIRCE project and the Moss Laboratory at the
University of Maryland), and we hope to
soon have a better understanding of
which physical features are important
for sculpting the sound spectrum.
Although the big brown bat head is considerably smaller than a human head,
the wavelengths of the frequencies it
uses are smaller as well. As a result, the
HRTF of this bat, at the frequencies of
interest, looks remarkably similar to
those of humans.
In humans, where the ears are far
enough apart to create as much as 1 ms
of interaural time difference (ITD), timing is used to estimate the arrival angle
at low frequencies. Echolocating bats,
in contrast, have small heads (less than
2 cm in diameter for the big brown
bat), and the distance between the two
ears of a typical insect-eating bat generates an interaural time difference of
less than 70 µs. While a few bats have
been shown to be sensitive to such tiny
interaural time differences, most bats
do not seem to utilize this cue. The
dominant cues for directional localization in echolocating bats appear to be
binaural comparisons of intensity and
monaural spectral cues.
To explore the use of these directional cues, we are modeling two different
types of neurons thought to be impor-
tant for localization: 1) a type that
receives binaural narrowband input and
responds to interaural-level differences
and 2) a type that receives monaural
broadband input and is selective to
spectral notches at different frequencies. Figure 3 shows our our narrowband sonar system, which produces
interaural intensity cues.
If we limit our sound directions to the
horizontal plane and analyze a single
frequency band, the difference in sound
level at the two ears [interaural level difference (ILD)] will be correlated with
different directions. In the bat, as in
other mammals, information about the
ILD is known to be coded at the earliest
stage of binaural processing by neurons
in the lateral superior olive (LSO). The
LSO receives excitatory inputs from the
ear on the same side of the brain and
inhibitory inputs from the ear on the
opposite side of the brain (Figure 4).
Roughly speaking, this means that LSO
neurons on the right side of the brain
respond to echoes coming from the
right side of the animal, and vice versa.
When the excitation created by an echo
is stronger than the inhibition to a particular LSO cell, the neuron will fire a
single spike for an echo. With an array of
cells with different thresholds, a range of
ILD thresholds is created. By observing
the response of this array of cells, the
brain can determine the sound direction. As each echo arrives, the array
responds to indicate echo direction.
We have designed a population of
LSO neurons in VLSI (a commercially
available, 1.5-µm CMOS process) and
have demonstrated direction-selective
responses with our sonar system [23].
This circuit is designed to operate in
the MOSFET subthreshold region of
operation, consuming only microwatts
of power.
Another important cue for sound localization utilizes the fact that the pinnas
produce a spectral notch, the frequen-
In midbrain areas specialized for
auditory processing (inferior colliculus),
neurons have been found that respond
to specific pulse-echo delays with relatively broad tuning. This type of population coding enables the representation
of multiple objects at different ranges,
provides a substrate for spatial memory,
and allows range-specific parameter
optimizations. For example, echo signals from faraway objects tend to be significantly lower in intensity and could
potentially be optimized for a lower
signal-to-noise ratio.
How does a neuron become tuned to
a particular pulse-echo delay (i.e.,
range)? Neurophysiological experiments have shown that these neurons
have slightly underdamped membrane
potential dynamics that provide a timing cue. During the outgoing sonar
pulse, these neurons are inhibited, thus
reducing the membrane potential
(Figure 5). As the membrane potential
reverts back to its resting state (each
neuron with its own time constant), it
momentarily overshoots its resting
state (“rebounds”); hence, a narrow
window of opportunity exists for a
temporally coincident excitation from
an arriving echo to push the potential
above its spiking threshold. Since these
range-sensitive neurons fire at the time
of the arriving echo, like the azimuthsensitive neurons described in the previous section, it is possible to “bind”
the two pieces of information together
using temporal coincidence.
To model these neurons, we
designed an array of low-power (∼ 500
µW) VLSI neurons (0.5 µm CMOS
process) with similar rebound dynamics
that respond to pulse-echo delays in the
1–30 ms range [26]. To model the
“binding” of range and azimuth, we
[FIG5] Range-tuned neurons are
differentially inhibited by the outgoing
pulse; if the echo arrives at the time of
the postinhibitory rebound, the neuron
will produce a spike.
Pulse-Echo Delay
Delay (Range)
cy of which varies moderately with
sound elevation and weakly with sound
azimuth. By detecting the notch frequency binaurally, a combined estimate of both azimuth and elevation
can be made. To model this technique
for localization, we are constructing
and testing an ultrasonic cochlea chip
and different pinnas shapes that produce direction-dependent frequency
notches like bat pinnas.
Echolocating bats specialize in highfrequency hearing using ultrasonic
sounds that have the most power in the
frequency range of 20–100 kHz [24].
While some bats are specialized for specific frequencies with very sharp
cochlear threshold-tuning curves
(Q10dB ∼ 400), we are studying bats
that use a broadband vocalization; they
are ultrasonic frequency generalists
(e.g., Myotis lucifugus) with thresholdtuning curves of modest sharpness
(Q10dB values in the range of 10–30
[24]) throughout the ultrasonic frequency range. Good frequency resolution is important for detecting and
estimating frequency notches.
To address this need, we have been
designing a binaural, ultrasonic filterbank chip using a gyrator-based filter
that achieves moderate quality-factor
(Q) bandpass filtering followed by spiking neurons for use in modeling the
bat echolocation system [25]. The filter
frequencies are spaced exponentially to
mimic the frequency spacing in the bat
cochlea. We have also designed a neuron circuit that encodes the echo
amplitude and produces minimal
spike-related power-supply noise (a
typical problem for mixed-mode circuits); it reports spike events using a
current-mode digital output. This binaural cochlea chip operates on about
500 µW of power.
Determining the distance to an object is
based on the time of flight of the sonar
pulse. At the speed of sound, the distance
to objects can be calculated by allowing
for approximately 17 cm for every millisecond of delay measured.
–2.34 –1.96 –1.56 –1.17 –0.78 –0.39 0 0.39 0.78 1.17 1.56 1.96 2.34
Spike Rate Difference
ILD (Azimuth)
[FIG6] By using the temporal coincidence of spikes from the LSO neurons and the delaytuned (range-tuned) neurons, 2-D receptive fields can be created.
[FIG7] The binaural sonar head shown
mounted on a Koala mobile robot (KTeam, Switzerland) is used to test
neurally plausible, multiple-object
collision avoidance algorithms based on
sonar data.
have recently designed and fabricated a
chip that detects the temporal coincidence between the firing of azimuthsensitive cells and these range cells to
create “two-dimensional cells” that
respond to specific combinations of
range and azimuth (see Figure 6).
How do we connect low-level computations and data representations to higher-level behavior such as collision
avoidance and insect capture? How do
animals represent the 3-D world to
determine where and how to move?
Mobile robot studies have struggled
with this question for decades, formulating many different types of approaches that ultimately depend heavily on
the sensory system used to collect data.
While precise mapping and trajectory
planning approaches dominated mobile
robot research in the early years, recent
efforts have been drawing upon new
understanding of an area of the mammalian brain known as the hippocampus, where cells are thought to rapidly
“learn” to signal when the animal is in
a particular location within a familiar
In our laboratory, we are currently
focused on the problem of rapid steering decisions that occur during insect
pursuit in unknown, cluttered environments. In pursuit problems, the sensory activity generated by a target can be
used in a straightforward manner to
determine the actions an animal should
take to capture the target (e.g., producing a turning rate in proportion to the
error angle). In collision avoidance,
however, sensory activity generated by
obstacles should only determine actions
the animal should not take to avoid collision. Given how little is known about
the neurobiology of spatial decision
making, can we design a neurally plausible algorithm that can successfully
combine both target pursuit and collision avoidance?
By representing space in terms of
emptiness or openness, we have begun
to create successful algorithms that
can combine information about both
goal directions and multiple obstacles
to make good choices for navigation.
An example software-based testbed that
has been constructed to explore these
ideas is shown in Figure 7. Constructed
by a summer undergraduate research
team, the “batmobile” can navigate
through a “forest” of cardboard tubes
without slowing down. We are currently developing a spiking, neuron-based
version of this algorithm in VLSI for
high-speed, two-dimensional navigation suitable for model cars and aircraft in the near future.
Beyond research and education, there
are many obvious commercial and
industrial applications of integrated
sensory systems implemented in lowpower VLSI. The development of a
small, sophisticated, power-efficient,
low-cost echolocation system has
many potential applications beyond
neural modeling. In the biomedical
realm, such devices are beginning to
be used as another option for collision
avoidance and spatial sensing for blind
or low-vision patients. These devices,
when properly scaled down, could also
be used to guide endoscopic instruments or provide additional information about distance to monocular,
visually guided surgical tools. Air-coupled sonar, as a basic sensor module
for mobile robotics, has not advanced
significantly beyond a narrow-beam,
closest-target sensor, despite decades
of use. With robotic vacuum cleaners
finally hitting the market, a low-power
module with significantly more sensing capability at low cost could facilitate a new range of commercial
products and toys that have the ability
to sense objects in the near-field like a
full set of whiskers.
From a micro-aerial vehicle perspective, while GPS has successfully enabled
long-range navigation, the final leg of
many desirable missions occurs in locations where the lack of GPS signals and
unmapped obstacles make navigation
untenable; such locations include inside
buildings, under the forest canopy, in
canyons, and in caves. Obtaining the
range to objects directly, while computing azimuth, sonar systems are a natural
complement to vision systems for these
challenging environments. When combined with an ornithopter airframe, a
nearly silent device (to humans), the
ability to fly in the darkness seems to be
within reach.
Overall, this project is proving to be
a wonderful framework in which to
pursue different types of scientific and
engineering-oriented research and education. Understanding bat echolocation
involves many interesting problems of
signal processing within the context of
biological data representations and
neural hardware.
Timothy K. Horiuchi received both the
B.S. degree in electrical engineering
(1989) and Ph.D. degree in computation and neural systems (1997) from
the California Institute of Technology.
He did postdoctoral work in the Zanvyl
Krieger Mind/Brain Institute at the
Johns Hopkins University and is now
an associate professor with a joint
appointment in the Electrical and
Computer Engineering Department
and the Institute for Systems Research
at the University of Maryland, College
Park (UMCP). He is also a member of
the Neurosciences and Cognitive
Sciences Program at UMCP. His
research interests are in computational neuroscience and the implementation of neural circuit architectures in
mixed-mode neuromorphic VLSIbased processors. He is currently pursuing the development of analog VLSI
chips that mimic the signal processing
and sensorimotor control of bat
echolocation. He is also involved in
efforts to improve the tools and techniques used in neurophysiology.
[1] J.D. Altringham, Bats: Biology and Behavior.
New York: Oxford Univ. Press, 1996.
[2] G. Neuweiler, The Biology of Bats. New York:
Oxford Univ. Press, 2000.
[3] C.A. Mead, Analog VLSI and Neural Systems.
Menlo Park, CA: Addison-Wesley, 1989.
[4] C. Mead, “Neuromorphic electronic systems,”
Proc. IEEE, vol. 78, no. 10, pp. 1629–1636, 1990.
[5] R. Douglas, M. Mahowald, and C. Mead,
“Neuromorphic analogue VLSI,” in Annual Reviews
in Neuroscience, vol. 18, W.M. Cowan, E.M.
Shooter, C.F. Stevens, and R.F. Thompson, Eds. Palo
Alto, CA: Annual Reviews Inc., 1995, pp. 255–281.
[6] M.A. Mahowald, “VLSI analogs of neuronal visual
processing: A synthesis of form and function,” Ph.D.
thesis, Dept. Computer Science, California Inst.
Technology, Pasadena, CA, 1992.
[7] K.A. Boahen, “Retinomorphic vision systems II:
Communication channel design,” presented at the
ISCAS Supplement, 1996.
[8] Z.K. Kalayjian and A.G. Andreou, “A silicon retina
for polarization contrast vision,” presented at the Int.
Joint Conf. Neural Networks, 1999.
[9] E. Fragniere, A. van Schaik, and E. Vittoz,
“Design of an analogue VLSI model of an active
Cochlea,” Analog Integrat. Circuits Signal
Processing, vol. 13, no. 1–2, pp. 19–35, 1997.
[10] J. Lazzaro and C. Mead, “Circuit models of sensory
transduction in the cochlea,” in Implementations of
Neural Networks, C. Mead and M. Ismail, Eds. Norwell,
MA: Kluwer Academic, 1989, pp. 85–101.
[11] R. Sarpeshkar, “Efficient precise computation
with noisy components: Extrapolating from an electronic cochlea to the brain,” Ph.D. thesis, California
Inst. Technology, Pasadena, CA, 1997.
for modeling intersegmental coordination,” presented
at the 17th Conf. Advanced Research in VLSI, 1997.
[18] T.K. Horiuchi and C. Koch, “Analog VLSI-based
modeling of the primate oculomotor system,” Neural
Comput., vol. 11, pp. 243–265, 1999.
[19] R.R. Harrison and C. Koch, “A robust analog
VLSI motion sensor based on the visual system of the
fly,” Auton. Robots, vol. 7, no. 3, pp. 211–224, 1999.
[20] S.-C. Liu and A. Usseglio-Viretta, “Visuo-motor
fly-like responses of a robot using a VLSI motionsensitive chips,” presented at Proc. Second Int. ICSC
Symp. Neural Computation NC’2000, Berlin,
Germany, 2000.
[12] L. Watts, “Cochlear mechanics: Analysis and
analog VLSI,” Ph.D. thesis, California Inst.
Technology, Pasadena, CA, 1993.
[21] V.A. Walker, H. Peremans, and J.C. Hallam, “One
tone, two ears, three dimensions: A robotic investigation of pinnae movements used by rhinolophid and
hipposiderid bats,” J. Acoust. Soc. Am., vol. 104,
no. 1, pp. 569–579, 1998.
[13] P.M. Furth and A.G. Andreou, “Cochlear models
implemented with linearized transconductors,” presented at ISCAS’96, 1996.
[22] R. Mueller, “A numerical study of the role of the
tragus in the big brown bat,” J. Acoust. Soc. Am.,
vol. 116, no. 6, pp. 3701–3712, 2004.
[14] T. Horiuchi and K.M. Hynna, “A VLSI-based
model of azimuthal echolocation in the big brown
bat,” Auton. Robots, vol. 11, no. 3, pp. 241–247, 2001.
[23] R.Z. Shi and T.K. Horiuchi, “A VLSI model of
the bat lateral superior olive for azimuthal echolocation,” presented at Intl. Symp. Circuits and Systems,
Vancouver, Canada, 2004.
[15] S.R. Deiss, R. Douglas, and A.M. Whatley, “A
pulse-coded communications infrastructure for neuromorphic systems,” in Pulsed Neural Networks, W.
Maass and C.M. Bishop, Eds. Cambridge, MA: MIT
Press, 1999, pp. 157–178.
[16] S. Fusi, M. Annunziato, D. Badoni, A. Salamon,
and D.J. Amit, “Spike-driven synaptic plasticity:
Theory, simulation, VLSI implementation,” Neural
Comput., vol. 12, pp. 2227–2258, 2000.
[17] S.P. DeWeerth, G.N. Patel, M.F. Simoni, D.E.
Schimmel, and R.L. Calabrese, “A VLSI architecture
[24] A.N. Popper and R.R. Fay, “Hearing by bats,” in
Springer Handbook of Auditory Research, vol. 5.
New York: Springer-Verlag, 1995.
[25] H. Abdalla and T.K. Horiuchi, “An ultrasonic filterbank with spiking neurons,” presented at the Int.
Symp. Circuits and Systems, Kobe, Japan, 2005.
[26] M. Cheely and T. Horiuchi, “A VLSI model of
range-tuned neurons in the bat echolocation system,” presented at the 2003 Int. Symp. Circuits and
Systems (ISCAS ‘03), 2003.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF