This full text paper was peer-reviewed at the direction of IEEE Instrumentation and Measurement Society prior to the

acceptance and publication.

Neuron Model Database for Arm-Based Multi-Core

Neuromorphic Computing
Bo Gong Jiang Wang Siyuan Chang
School of Electrical and Information School of Electrical and Information School of Electrical and Information
Engineering Engineering Engineering
Tianjin University Tianjin University Tianjin University
Tianjin, China Tianjin, China Tianjin, China
gongbo@[Link] jiangwang@[Link] changsiyuan@[Link]
2023 IEEE International Instrumentation and Measurement Technology Conference (I2MTC) | 978-1-6654-5383-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/I2MTC53148.2023.10176069

Weitong Liu Jixuan Wang Xile Wei

School of Electrical and Information School of Electrical and Information School of Electrical and Information
Engineering Engineering Engineering
Tianjin University Tianjin University Tianjin University
Tianjin, China Tianjin, China Tianjin, China
liuweitong@[Link] wangjixuan@[Link] xilewei@[Link]

Abstract—The development of neuromorphic computing engineering projects, the simulation platform basically
provides new approaches for research in neuroscience. However, consists of three parts: models, the hardware design, and the
the network construction phase occupies most of development human-machine interface.
cycle. Therefore, this paper builds a multi-scale neuron model
database for the developed ARM-based multi-core In recent years, with the integration of neuronal dynamics,
neuromorphic computing system, which is perfect matched to biomedicine and computer science, the construction of virtual
the embedded system from ion channel and neuron scales. By brain is playing an increasingly significant role for
optimizing the calculation process of neuron models, it can biomedicine [7] [8]. As the major part of the central nervous
support the real-time simulation of various models in the system, the brain is the most typically unstable and
database, which provides strong support for the simulation of indeterminate system whose most basic structural and
the system in large-scale brain networks and spiking neural functional units are neurons [9]. There are many types of
networks (SNNs). The Leaky Integrate and Fire (LIF), neurons and each has its own unique function. However, it is
Izhikevich, and Hodgkin-Huxley (HH) neurons are simulated in indisputable that neurons basically output nerve impulses by
the system within 0.5s, 0.1s, and 0.5s, respectively, which depolarizing to produce action potentials [10], which is the
verified the great potential of the neuron model database in SNN basis for human beings to be able to think, memorize, and feel.
modeling, large-scale brain network simulation and The functions that can be performed by individual neurons are
neurodynamic analysis. The neuron model database serves as
very limited, but networks composed of a large number of
the foundation for network construction, and study upon this
will provide a normalized method for embedded neuromorphic
neurons can perform unimaginably complex functions.
computation. Therefore, how to select the appropriate model among the
numerous neuron models to construct local networks of the
Keywords—neuron model database, neuromorphic computing, brain is still one of the challenges in neuroscience.
Leaky Integrate and Fire (LIF) model, Izhikevich model, In the construction of large-scale neuronal networks, it is
Hodgkin-Huxley (HH) model necessary to determine the type and the number of neurons,
I. INTRODUCTION the way of synaptic connections, the network model, etc.
Currently available simulators such as NEURON [11], NEST
With the development of neuromorphic engineering, more [12], etc. They are relatively complete solutions proposed for
and more computational platforms have emerged, which local computers, but they may not be applicable to embedded
provide brand new methods and approaches for research in neuromorphic system. Moreover, neuromorphic engineering
neuroscience. Neuromorphic computing has not yet formed a requires different scales of models such as ion channel models,
systematic theoretical framework. Currently, there are a spiking neuron models, and cluster models. To address the
number of popular approaches: digital circuits (IBM's above needs, we have built a multi-scale personalized neuron
TrueNorth project) [1], customized chips (Intel's Loihi project) model database for the constructed system, which is
[2], application specific integrated circuits (ASICs) completely developed in C/C++ language with excellent
(BrainScales project at Heidelberg University) [3], massively reliability and scalability.
parallel computing (SpiNNaker project at the University of
Manchester) [4] and memristors such as the Tianjic project at This paper is composed of four themed sections. The 2nd
Tsinghua University [5]. There are still many gaps in the section of the paper describes the neuron and ion channel
cognitive mechanisms of the brain, and the functioning of the models in the model database and matches them with the
biological brain has not been fully unraveled. Therefore, developed embedded platform. The 3rd section tests the
research on neuromorphic engineering is urgent and model database and collects the data generated in the platform
indispensable. We have designed an ARM-based multi-core for analysis. The 4th section concludes the paper.
embedded neuromorphic system [6], which has significant
advantages in terms of power consumption, computational II. NEURON MODEL DATABASE
efficiency and scalability. Consistent with most neuromorphic In embedded neuromorphic computing, the neuron model
is the most fundamental unit in the database, which modularly
This work was supported by the National Natural Science Foundation of stores varies types of neurons required in different simulations.
China (Grant No.62071324 and 62171312). Corresponding author: Xile
Wei.

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 22,2024 at 04:42:30 UTC from IEEE Xplore. Restrictions apply.
 Therefore, this section will cover the models and their TABLE I. PARAMETERS OF VARIOUS MODELS
applications in the neuromorphic computing platform. Parameters HH Model RTM Model Ersir Model
A. Neuron Models 𝐶 1 𝜇𝐹/𝑐𝑚 1 𝜇𝐹/𝑐𝑚 1 𝜇𝐹/𝑐𝑚
The features and phenomena that need to be observed in 𝐸 50 mV 50 mV 60 mV
various local networks of the brain are different. In particular, 𝐸 -77 mV -100 mV -90 mV
when building large-scale networks, using complex neuron
𝐸 -54.4 mV -67 mV -70 mV
models will increase the simulation time and computational
stress of the network significantly. Therefore, the most basic 𝑔̅ 120 𝑚𝑆/𝑐𝑚 100 𝑚𝑆/𝑐𝑚 112 𝑚𝑆/𝑐𝑚
and essential element of the model database is the wide range 𝑔̅ 36 𝑚𝑆/𝑐𝑚 80 𝑚𝑆/𝑐𝑚 224 𝑚𝑆/𝑐𝑚
of neurodynamic models. In 1952, Hodgkin and Huxley 𝑔̅ 0.3 𝑚𝑆/𝑐𝑚 0.1 𝑚𝑆/𝑐𝑚 0.5 𝑚𝑆/𝑐𝑚
proposed the Hodgkin-Huxley (HH) model [13] by analyzing
.
the stimulation potential data of squid nerves, which was the 𝛼
⎧ . / .
first to use ion-gated channels to describe the effect of
⎪
different ion concentrations inside and outside the membrane ⎪ 𝛼
on the membrane potential of neurons , and explained the ⎪ / .

process of action potential generation and propagation. Since Ersir Model: 𝛼 0.0035exp 𝑉/24.186 (5)
then, more and more neuron models have been developed to ⎨𝛽 1.2262exp 𝑉/42.248
describe the dynamic properties of neurons. Along with HH ⎪𝛽 0.025exp 𝑉/22.222
⎪
model, Leaky Integrate and Fire (LIF) [14] and Izhikevich ⎪ . .
models [15] are also commonly used in network construction, ⎩ 𝛽 . / .
and the model database is mainly built based on these three Currently, the LIF neuron model is more frequently used
models and their variants. in Spiking Neural Networks (SNN). It ignores the specific
The general expression of the HH model is given by kinetics of neuron action potential , only considers the passive
properties of membrane potential and the triggering of action
𝐶 𝑔̅ 𝑚 ℎ 𝐸 𝑉 𝑔̅ 𝑛 𝐸 𝑉 𝑔̅ 𝐸 𝑉 𝐼  potential. This type of model is easier to describe the
mechanism of information transmission by encoding in brain
𝛼 1 𝑋 𝛽 𝑋 𝑋 ∈ 𝑚, 𝑛, ℎ   networks. The specific form is as follows:
where 𝑉 is the membrane potential of the neuron, 𝐶 is the 𝐶 𝑔 𝑉 𝐸L 𝐼 𝑡 (6)
membrane capacitance, 𝐼 represents the total transmembrane
current, and the equilibrium potentials of 𝑁𝑎 , 𝐾 and leaky 𝑖𝑓 𝑉 𝑉 , 𝑡ℎ𝑒𝑛 𝑉 𝑉rest (7)
channels are represented by 𝐸 , 𝐸 , 𝐸 . 𝑔̅ , 𝑔̅ and 𝑔̅ are When the membrane potential 𝑉 reaches the threshold 𝑉 ,
the maximum conductance of each ion channel, 𝑚, ℎ, and 𝑛 the neuron generates an action potential and then returns to the
are the gating variables of the ion channels, 𝛼 and 𝛽 are the resting potential 𝑉rest at the next moment. c, g and 𝐼 𝑡 are the
rate constants. capacitance, conductance and external input current of the
Besides, there are some variants of the HH model, such as membrane, respectively.
the Reduced Traub-Miles (RTM) model, which only changes Izhikevich model is a simplified neuron model that
the rate constant of the HH model; and the Ersir model, which achieves a balance between computational complexity and
modifies the current of the potassium ion channel in the HH dynamic performance between HH and LIF models. With
model to 𝐼 𝑔̅ 𝑛 𝑉 𝐸 [16]. The common parameters different parameters, Izhikevich neurons are able to generate
of these models are given in Table I. different frequencies of spikes, thus forming different firing
In the above mentioned models, the rate constants follow: rhythms in the plastic network, which makes it ideally suited
for the simulation of large networks. The specific form is as
. follows:
⎧ 𝛼
/
⎪ . 0.04𝑉 5𝑉 140 𝑢 𝐼 (8)
⎪ 𝛼
⎪ /

HH Model: 𝛼 0.07exp 𝑉 65 /20 (3) 𝑎 𝑏𝑉 𝑢 (9)

⎨ 𝛽 4exp 𝑉 65 /18
⎪𝛽 With the resetting rule:
⎪ 0.125exp 𝑉 65 /80
⎪ 𝑉←𝑐
𝛽 𝑖𝑓 𝑉 30𝑚𝑉, 𝑡ℎ𝑒𝑛 (10)
⎩ / 𝑢←𝑢 𝑑
𝛼
. The time constant of the recovery variable 𝑢 is parameter
⎧ / 𝑎 , which is used for the slow recovery of the membrane
⎪ . potential 𝑉. The parameter 𝑏 represents the dependence of 𝑢
⎪ 𝛼
/ on 𝑉. The value of the after-spike resetting is determined by
⎪
𝛼 0.128exp 𝑉 50 /18 the parameter 𝑐 , and 𝑢 is increased by 𝑑 . By varying the
RTM Model: (4)
⎨ 𝛽
. values of the four dimensionless parameters 𝑎, 𝑏, 𝑐, 𝑑 ,
/ Izhikevich model is able to perform a variety of neuron firing
⎪
⎪ 𝛽 0.5exp 𝑉 57 /40 patterns. Typically, the parameter values are given in Table 2.
⎪
⎩ 𝛽
/

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 22,2024 at 04:42:30 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. PARAMETERS OF IZHIKEVICH MODEL

Parameters a b c d
Values 0.02 0.20 -55.0 4.00

B. Ion Channel Models

Electrophysiological activity in neurons is maintained and
transmitted primarily by ionic currents passing through the
neuron membrane. Most of these transmembrane currents
involve several of the four ions: 𝑁𝑎 , 𝐾 , 𝐶𝑎 and 𝐶𝑙 . The
concentration of these ions varies both inside and outside the
cell, resulting in an electrochemical gradient, which is the
main force driving neural activity.
Two dominant forces, concentration and potential
gradients, will drive each ion through the channel, which is
formed by large protein molecules across the cell membrane.
The ions diffuse along a concentration gradient, for example,
𝐾 diffuses outside the cell because the concentration of 𝐾 Fig. 1. Physical diagram of the BCU. The BCU is composed of 9
inside the cell is higher. When leaving the cell, 𝐾 carries STM32F4 series microprocessors, equipped with a rich set of external
positive charge and leaves negative charge inside the cell, expansion interfaces.
resulting in an outward current. The positive and negative
charges accumulate on either side of the membrane surface, surrounding 8 microprocessors. The Static Random Access
forming a membrane voltage. This potential slows the Memory (SRAM) and Read-Only Memory (ROM) in the
diffusion of potassium ions as it is attracted to the negatively BCU are accessible to all processors and the data generated in
charged interior of the membrane and repelled by the each processor is recorded in a time-division multiplexed
positively charged exterior part. When the forces applied by manner. Thus, the neuron model database is integrated in the
the concentration and potential gradients balance each other, chip's Flash and the relevant parameters are saved in the ROM
the total transmembrane current is zero. for recall at any time.

V represents the membrane potential, 𝐸 , 𝐸 , 𝐸 and In this platform, considering the high computational real-
𝐸 represent the equilibrium potential. When membrane time requirements of the neuron dynamics network model in
potential is equal to equilibrium potential, 𝐼 0. Otherwise, specific application scenarios, it is necessary to ensure that the
simulation is consistent with the real time scale. We found that
the 𝐾 current is proportional to the potential difference:
it is difficult to realize real-time simulation by using standard
𝐼 𝑔 𝑉 𝐸   function libraries due to the complexity of the model and the
limitation of the chip computing capability. A more detailed
𝑔 is the conductance of 𝐾 , and other major ion currents can analysis of the time overhead of each process in the model
be expressed in this form, thus the general ion channel current calculation shows that the most time-consuming calculation
will be: process is focused on the operation and update of gating and
𝐼 𝑔̅ 𝑚 ℎ 𝑉 𝐸 𝐴  intermediate variables, and the results are shown in Fig. 2.
𝑔̅ is the maximum conductance, 𝐸 is the inversion Furthermore, the results of the simulation debugging show
potential of the ion channels. 𝐴 is the surface area of that the key reason affecting the real-time performance is the
compartment 𝑖 containing these ion channels. 𝑚 and ℎ excessive number of floating-point and exponential operations
are the activation and deactivation gating variables with the in the model, which leads to poor computational efficiency.
following expressions: Therefore, the model is reasonably simplified for this process,
and the calculation of the exponential function is replaced as
𝜏 𝑚 𝑚  follows:

𝜏 ℎ ℎ 
𝜏 and 𝜏 represent the time constants of the activation
and deactivation variables, 𝑚 and ℎ are the corresponding
steady-state values. (15) (16) (17) and (18) are all functions of
the membrane potential V.
C. Implementation on Multi-Core Embedded Platform
We have developed a multi-core distributed neuromorphic
computing platform. With multi-node decentralized parallel
computing, it can meet the computational and memory
requirements of large-scale networks. Moreover, the two-
dimensional grid-like topology of the system allows flexible
scaling of basic community units (BCUs) (as shown in Fig. 1) ,
which enables applications in multiple scenarios and tasks. Fig. 2. The time overhead ratio of each process in the iterative solution
The microprocessor in the central position of the BCU acts of the model. Calculation and updating of gating and intermediate
variables take up more than 60% of the calculation time.
as the processing center and is fully connected to the

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 22,2024 at 04:42:30 UTC from IEEE Xplore. Restrictions apply.
 𝑒 𝑙𝑖𝑚 1 𝑥/𝑛 (15)
→

Thus, 𝛼 in (3) is simplified by:

𝛼 0.07 𝑙𝑖𝑚 1 𝑉 65 /20 /𝑛 (16)
→

This approximation increases the accuracy as 𝑛 increases,

while the computational efficiency decreases. To minimize
the number of calculations, 𝑛 is converted to an exponential
multiple of 2. This method will be used in solving the
difference equation, where (16) is replaced by:

𝛼 ≔1 /2
𝛼 ≔𝛼 ∗𝛼
⋮ 𝑘 times (17)
Fig. 4. Simulation of LIF neuron. The membrane potentials of LIF
𝛼 ≔𝛼 ∗𝛼 neurons were simulated when external currents 𝐼
𝛼 ≔ 0.07 ∗ 𝛼 0, 2, 4, 5, 6, 𝑎𝑛𝑑 9𝑛𝐴, respectively.
By switching on the floating-point unit (FPU) and this
conversion, as well as using a single-precision floating-point the acquisition board for further processing, as shown in Fig.
format, the time overhead of model computation is greatly 4. It can be found that the LIF neuron fires constantly when
reduced and the real-time performance of the system is the external input current reaches 5𝑛𝐴. And the frequency of
improved. One of the significant changes is that the gating LIF neuron spikes increases as external input current goes up.
variable and intermediate variable computation percentages
are reduced to 41.91% versus 19.77%. This method can be Similarly, the Izhikevich neuron model was simulated in
applied to the calculation of a variety of neural dynamics the platform with iteration steps as 0.1 𝑚𝑠 and simulation
models, which is beneficial to the subsequent functional time as 100 𝑚𝑠 . We obtained a more accurate membrane
development of the neuromorphic computing platform. potential with the Runge-Kutta method, as shown in Fig. 5.

III. EXPERIMENTS AND RESULTS

In the experiments, we simulated the LIF, Izhikevich, HH
models and its variants by a BCU. The experiment is shown
in Fig. 3. During the simulation, the model database is
integrated in the Flash of the chip, and the neuron membrane
potentials generated in the BCU will be sent to the acquisition
board through the AD/DA port. And then, the data in the
acquisition board is packaged and sent to the host PC through
the serial interface, and oscilloscope will display the current
changes of the neuron membrane potential.
We simulated the LIF neuron model in the platform, with
setting the iteration step to 0.01 𝑚𝑠 and the simulation time to
0.5 𝑠. The spikes of the LIF neuron under different external
constant currents were calculation by the forward Euler
method, and the data were uploaded to the host PC through

Fig. 5. Several firing patterns of Izhikevich neurons. (a) Regular

spiking (RS). (b) Resonator (RZ). (c) Thalamo-cortical (TC). (d) Low-
Fig. 3. Experimental diagram. The overall experiment consists of the
threshold spiking (LTS). (e) Intrinsically bursting (IB). (f) Fast spiking
host PC, oscilloscope, BCU group, acquisition board, and serial
(FS). (g) Chattering (CH).
interface.

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 22,2024 at 04:42:30 UTC from IEEE Xplore. Restrictions apply.
 Izhikevich and HH models are simulated in 0.5s, 0.1s and 0.5s,
respectively. Moreover, the forward Euler and Runge-Kutta
methods are applied. It is proved that the LIF neuron model is
highly simplified and has the lowest calculation complexity,
which is more suitable for the application in SNN modeling.
The Izhikevich neuron model is able to reproduce various
firing states, which has considerable advantages in the
simulation of large-scale brain networks. HH model and its
variants described the multiple properties of neurons in detail
from the ion channel scale, which also leads to the highest
calculation complexity. In this paper, the calculation of the
HH model are optimized, but it is still not suitable for large-
scale brain network simulations.
Efficient implementation of artificial neural networks
(ANNs) in an embedded neuromorphic computing platform
Fig. 6. Simulation of the HH model and its variants when 𝐼 0.39𝜇𝐴. may be a challenge, and the usual approach is to transform
them into SNNs before implementation. A version of MNIST
containing 50K training and 1K test samples with a network
size of 320 neurons, deployed in 4 BCUs, has been
implemented in this platform so far [6]. The synaptic weights
are segmented and saved in the ROM of the corresponding
BCUs, and the MNIST images are encoded as a spike train,
which is presented to each BCU via the routing bus, enabling
lightweight porting of the network. The neuron model
database is built mainly to serve the constructed 32-bit ARM
embedded neuromorphic computing platform and provides a
normalized approach to it. However, in 64-bit systems, the
algorithm still needs to be further improved. Future work will
continue to extend and improve this database, which lays a
solid foundation for embedded neuromorphic computing in
large-scale brain networks, SNNs and ANNs.
REFERENCES

Fig. 7. Simulation of the HH model and its variants when 𝐼 0.68𝜇𝐴. [1] F. Akopyan, J. Sawada, A. Cassidy, et al., “TrueNorth: Design and
tool flow of a 65 mW 1 million neuron programmable neurosynaptic
With changing the values of 𝑎, 𝑏, 𝑐, and 𝑑 in (9), (10), the chip,” IEEE Trans. Comput. Aided Des. Integr. Circuits Syst., vol. 34,
Izhikevich model can simulate a variety of neuron firing forms. pp. 1537-1557, 2015.
Compared with the LIF model, it adds more descriptions of [2] M. E. Davies, N. Srinivasa, T. Lin, et al., “Loihi: A neuromorphic
neuronal properties while sacrificing some computational manycore processor with on-chip learning,” IEEE Micro, vol. 38, pp.
82-99, 2018.
efficiency, which is important for large-scale brain network
[3] S. Schmitt, J. Klaehn, G. Bellec, et al., “Neuromorphic hardware in
simulation. the loop: Training a deep spiking network on the BrainScaleS wafer-
The HH, RTM, and Ersir models were also simulated in scale system,” 2017 Int. Joint Conf. on Neural Networks (IJCNN), pp.
2227-2234, 2017.
the platform with iteration steps as 0.01 𝑚𝑠 and simulation
[4] S. Furber, F. Galluppi, S. Temple and L. Plana, “The SpiNNaker
time as 0.5 𝑠. When the external input current was 0.39 𝜇𝐴, project,” Proc. IEEE, vol. 102, pp. 652-665, 2014.
the HH and Ersir models were unable to fire constantly, as [5] J. Pei, L. Deng, S. Song, et al., “Towards artificial general intelligence
shown in Fig. 6. When the external input current is 0.68𝜇𝐴, with hybrid Tianjic chip architecture,” Nature, vol. 572, pp. 106-111,
the frequency of the RTM model is significantly higher than 2019.
that of the HH and Ersir models, as shown in Fig. 7. [6] X. Wei, J. Xu, B. Gong, et al., “Multi-core ARM-based hardware-
accelerated computation for spiking neural networks,” IEEE Trans.
With the optimization, the most complex HH model can Ind. Inf., 2022, doi: 10.1109/TII.2022.3216011.
also be simulated real-time in the platform, which shows the [7] Z. Wan, Y. Dong, Z. Yu, et al., “Semi-supervised support vector
rich dynamics of neurons from the ion scale. Comparing Fig. machine for digital twins based brain image fusion,” Front. Neurosci.,
6(b) with Fig. 7(b), the frequency of the RTM model also vol. 15, pp. 705323, 2021.
increases as the external input current goes up. [8] M. Schirner, L. Domide, D. Perdikis, et al., “Brain simulation as a
cloud service: The Virtual Brain on EBRAINS,” NeuroImage, vol:
IV. CONCLUSION 251, pp. 118973, 2022.
[9] M. Breakspear, “Dynamic models of large-scale brain activity,” Nat.
This paper proposed a neuron model database for ARM- Neurosci.,vol. 20, pp. 340-352, 2017.
based multi-core neuromorphic computing, which contains [10] A. Gidon, T. Zolnik, P. Fidzinski, et al., “Dendritic action potentials
many types of neuron and ion channel models with great and computation in human layer 2/3 cortical neurons,” Science, vol:
significance for large-scale brain network simulation and 367, pp. 83–87, 2020.
SNN modeling. In order to adapt the model database perfectly [11] N. Carnevale, “The NEURON simulation environment,” Neural
to the developed embedded platform, we have optimized the Comput., vol. 9, pp. 1179-1209, 1997.
model calculation steps so that the platform can simulate a [12] M. Gewaltig and M. Diesmann, “NEST (NEural simulation tool),”
wide range of neurons in real-time. In the experiments, LIF, Encycl. Comput. Neurosci., vol. 2, pp. 1430, 2007.

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 22,2024 at 04:42:30 UTC from IEEE Xplore. Restrictions apply.
 [13] A. Hodgkin and A. Huxley, “A quantitative description of membrane [15] E. Izhikevich, “Simple model of spiking neurons,” IEEE Trans.
current and its application to conduction and excitation in nerve,” J. Neural Networks, vol. 14, pp. 1569-72, 2003.
Physiol., vol. 117, pp. 500-44, 1952. [16] C. Börgers, Three simple models of neurons in rodent brains. In: An
[14] A. Burkitt, “A review of the integrate-and-fire neuron model: I. introduction to modeling neuronal dynamics. Texts in Applied
Homogeneous synaptic input,” Biol. Cybern., vol. 95, pp. 1-19, 2006. Mathematics, vol. 66, Springer, Cham, 2017.

Authorized licensed use limited to: Visvesvaraya Technological University Belagavi. Downloaded on March 22,2024 at 04:42:30 UTC from IEEE Xplore. Restrictions apply.