IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 65, NO.

5, MAY 2018 577

Development and Applications of Biomimetic

Neuronal Networks Toward BrainMorphic
Artificial Intelligence
Timothée Levi , Takuya Nanami, Atsuya Tange, Kazuyuki Aihara, and Takashi Kohno

Abstract—This brief presents the brainmorphic artificial 2) designing new low-power neuromimetic systems for AI
intelligence (BMAI) project. The main goal is to generate a novel applications.
type of neuromorphic computing systems including novel algo- The biomimetic neural network is a neuromorphic system
rithms and devices, which can achieve very high performance of with a most detailed level of analogy to the nervous system. It
artificial intelligence (AI) information processing with low power
consumption and which can be very close to biological systems. is a network of silicon neurons connected via silicon synapses
To reach this goal, we developed biomimetic neural networks. and plasticity rules. It can be silicon based [1] or microfluidic
Two systems have been designed, one analog chip and one digi- based [2]. To present the different hardware implementations
tal system into FPGA. We present in this brief two applications of of these systems [3], we rely on the review of neuromorphic
these biomimetic neural networks, one in AI with a pattern recog- architectures published in 2010, [4] and we enrich it with other
nition algorithm and one in neuroscience for understanding the articles. It will allow us to show the growing interest in neu-
hearing system of Drosophila for designing a new asynchronous romorphic systems these past/latest years, and their utilities
low-power sensor system in the future.
and their potential. Indeed, their applications extend to high-
Index Terms—Artificial intelligence, biomimetic neural energy physics, image processing (object / image / pattern
network, neuromorphic engineering, silicon neuron. recognition, image segmentation, video processing), robotics,
acoustic or olfactory recognition and understanding of the ner-
vous system. Hardware implementations are divided into two
I. I NTRODUCTION major categories: analog implementation (based on dedicated
EUROMORPHIC systems are designed by mimicking chips) and digital implementation (based on FPGA, micro-
N or being inspired by the nervous system, which real-
izes robust, autonomous, and power-efficient information pro-
processors, microcontrollers or neurochips). We focus here on
biomimetic and non-bio-inspired systems. The very first plat-
cessing by highly parallel architecture. The purpose of the forms appeared around twenty years ago [5]–[7]. In the case
Brainmorphic artificial intelligence project (BMAI), is to gen- of analog implementations, some systems implement multi-
erate a novel type of neuromorphic computing systems includ- compartmental models [8], conductance models [9] or with
ing novel algorithms and devices, which can achieve very threshold models [10], [11]. These platforms start from an
high performance of Artificial Intelligence (AI) information analog computing core, usually an ASIC, which describes the
processing with low power consumption and close resem- electrophysiological activity of the neuron. The architecture of
blance to biological systems. Using biomimetic neural network the different platforms results from a compromise between the
systems, we construct a novel fundamental technology of computational cost and the complexity of the model (directly
the information processing by developing the Brainmorphic correlated to biological plausibility). The integration of plas-
AI system which can execute fast autonomous informa- ticity and synapses is usually done by a digital map that makes
tion processing with low energy. Two main objectives are the link between different analog chips. As for the digital
raised: 1) understanding and simulating biological intelligence; implementation, biomimetic neural network implementations
on FPGAs are generally used for pattern recognition, image
Manuscript received February 26, 2018; revised April 5, 2018; accepted
segmentation, video image processing and video analysis [12].
April 5, 2018. Date of publication April 9, 2018; date of current version The last point of the discussion concerns the size of artifi-
May 1, 2018. This work was supported in part by NEC Corporation, in part cial networks. Indeed, neurons are organized in networks of
by the JST PRESTO program, WPI, MEXT Japan, in part by the JST CREST various sizes. Biological plausibility is a constraint on the max-
program, and in part by JSPS KAKENHI under Grant 15H05707 and Grant
17K19450. This brief was recommended by Associate Editor J. M. de la Rosa. imum possible size of the network. To obtain large networks
(Corresponding author: Timothée Levi.) of neurons (population of 1000 neurons or more), it is
T. Levi, T. Nanami, A. Tange, and T. Kohno are with the necessary to:
Institute of Industrial Science, University of Tokyo, Tokyo 153-8505, • implement a simple neural model and a simple synapse
Japan (e-mail: [email protected]; [email protected]; tange@
sat.t.u-tokyo.ac.jp; [email protected]). model. Reference [13] presents an implementation on an ASIC
K. Aihara is with the Institute of Industrial Science, University of Tokyo, containing one million LIF (Leak Integrate and Fire) neurons
Tokyo 153-8505, Japan, and also with the WPI-IRCN, UTIAS, University of and 256 million synapses.
Tokyo, Tokyo 153-8505, Japan (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available
• perform accelerated neural network calculations such as
online at http://ieeexplore.ieee.org. the SpiNNaker [14] platform with a multiprocessor architec-
Digital Object Identifier 10.1109/TCSII.2018.2824827 ture. With over one million cores, and one thousand simulated
1549-7747 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
578 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 65, NO. 5, MAY 2018

neurons per core, the machine will be capable of simulating

one billion neurons. We also have the BrainScaleS system
which is a stack of modules composed of 20 wafers integrating
448 neuromorphic chips and a routing system. Each module
emulates the activity of 512 neurons and 115,000 synapses,
performing calculations 10 000 times faster than biological
time [15]. However, all these platforms are not working in bio-
logical real-time but accelerated time and then are not a good
option to perform bio-hybrid experiments. Another main draw-
back is that the neuron model is bioplausible but these systems
do not allow axonal delay, biomimetic plasticity and synap-
tic noise [16]. All of these different points are mandatory for
reproducing the biological dynamics of living neural networks.
With the emergence of real-time neuromorphic platforms,
significance of bio-hybrid experiments is increasing not only
for the development of neuromorphic biomedical devices [3],
but also for elucidating the mechanisms of the information Fig. 1. Block diagram of our analog VLSI silicon neuron. The spiking
dynamics is produced by the fast subsystem (v- and n-blocks). The q-block
processing in the nervous system by the approach of “analysis produces slow dynamics that modulates it. The voltage clamp amplifier (top
by synthesis”. Reference [17] introduces the term “neurobio- block) provides a means to draw the v-, n-, and q-nullclines experimentally.
hybrid” for one system formed by at least living neurones gx (v) and rx (x) circuits generate a shallower sigmoidal curve. fx (v) circuits
and at least one artificial entity that establish uni or bi- generate a steeper sigmoidal curve.
directional communications between them. The characteristics
of the systems for neurobiohybrid experiments are biolog-
ical real-time, complex neuron models and plasticity that on the timing of spike generation but variability of spike inten-
reproduce temporal neuronal activity and / or action poten- sity and phase response curve (PRC) are also important in
tial morphology. References [18]–[20] present different works brain signal processing. That is one reason for developing
on closed-loop hybrid experiment [21], which is characterized a qualitative silicon neuron model [31]. It is a good trade-off
by the interconnection between biological neural networks between biological plausibility and low resource implementa-
and their artificial counterparts [22]. Neurobiohybrid systems tion. It can mimic several classes of neuronal activities that
are more and more studied [23] and even a geographically cannot be realized by LIF and IZH models. Its equations are:
distributed bio-hybrid neural network is now possible [24]. dv
C = fv (v) − gv (v) − rn (n) − rq (q) + Iav + IStim (1)
To conclude, there is a trade-off between the bio-plausibility dt
of the model and simplicity and power-consumption of the dn
circuit. Power-efficient and simple silicon neurons assume = fn (v) − gn (v) + Ian − rn (n) (2)
dt
uniform spikes, but biophysical experimental data [25], [30] dq
suggest the possibility that variety of spikes given to a synapse = fq (v) + Iaq − rq (q) (3)
dt
is playing a certain role in the information processing in
the brain. Reference [26] indicates that subthreshold varia- where v is the membrane potential, n the variable for fast
tion in the presynaptic membrane potential also determines dynamics and q the variable for slow dynamics. Ia represents
spike-evoked transmission and that membrane voltage might a constant leak current. f, g and r functions are monotonic
modulate neurotransmitter release. Reference [27] discusses increasing sigmoidal functions of the idealized V–I charac-
the possible real neural coding used in the brain. We think teristic curves of the differential pair and transconductance
that being closer to the biology will lead to better results and amplifier.
allow neurobiohybrid experiments that bio-inspired systems
cannot perform. B. Analog Silicon Neuron on Chip
We have designed an analog ASIC with TSMC 0.25µm
technology [32], [33] that can realize several classes of
neuronal activities including Fast Spiking (FS) with both
II. B IOMIMETIC N EURAL N ETWORKS
Class I and II in the Hodgkin’s classification [34], Regular
A. Qualitative Silicon Neuronal Modeling Spiking (RS), Square-Wave Bursting (SWB) and Elliptic
Silicon neuron and synapse models for the Brainmorphic Bursting (EB). Fig. 1 describes the block diagram of this
information processing systems have to be designed care- system.
fully so that they inherit the features in the neuronal activity To design the chip, matured CMOS circuitries are used:
as authentically as possible. In many neuromorphic circuits differential pair, transconductance amplifier, etc. Field-effect
with simple circuitry, neuronal spikes are assumed to always transistors (FETs) are operated in their subthreshold domain
have the same shape and magnitude in the I&F-based mod- for low power consumption. To demonstrate the efficiency of
els including the Izhikevich model (IZH) [28] and LIF our circuit, we describe Elliptic bursting mode which is one
model [29], although biophysical experiment results suggest of the most complex to implement. Conventional silicon neu-
the possibility that the variation in the spikes can be playing rons cannot reproduce this rhythmic activity. Elliptic bursting
important roles in the information processing in the nervous is a class of bursting activity where a Class II spiking system
system [30]. Conventional silicon neuron models only focus has slow negative feedback dynamics. The fast subsystem of
LEVI et al.: DEVELOPMENT AND APPLICATIONS OF BIOMIMETIC NEURONAL NETWORKS TOWARDS BMAI 579

Fig. 2. Experimental result of the analog chip for the elliptic bursting mode.
The yellow curve is the membrane voltage v and the green curve represents
signal q. Time scale is 200ms by division, and voltage scale is 50mV by
division.

Fig. 4. Architecture of a neuronal network listening incoming spike trains

with hidden patterns. Wji are the synaptic weigh from neuron j (pre-synaptic)
to neuron I (post-synaptic). Inhibitory synapses are represented by white
circles and excitatory ones by black circles.

Fig. 3. Experimental results of A) RS neuron and B) LTS neuron. Scope III. A PPLICATIONS
pictures after 12-bit DAC module using SPI communication. Time scale is
200ms by division for A) and 10ms for B), and voltage scale is 20mV by Here, we present two applications for our biomimetic silicon
division for both pictures. neural network. One for AI and one for neuroscience. Both
will be tested first on digital circuits and then implemented in
analog chips.
our silicon neuron can realize Class II and we could realize the
elliptic bursting experimentally as shown in Fig. 2. 16 param- A. Pattern Recognition
eter voltages have to be tuned for intended neuronal activity. Spike-Timing-Dependent Plasticity (STDP) is a biological
Our analog silicon neuron circuit consumes only less than process that adjusts synaptic efficacy between neurons in the
5 nW. For every neuron mode with typical firing rate (less brain. The process adjusts synaptic efficacy based on the rel-
than 100Hz), the power consumption did not exceed 5nW. It ative timing of a particular neuron’s output and input action
is in the same range as [11], [35], and [36] but can repro- potentials or spikes. Recently, it has been shown how STDP
duce more classes of neuronal activities. This ASIC could could play a key role in detecting repeating patterns and in
be used for neurobiohybrid experiments. As geographically generating selective response to them [38]. The concept of
distributed artificial biomimetic neural networks and living STDP has been shown to be an important learning algorithm
neurons already exist [23], using this ASIC in combination for forward-connected artificial neural networks in pattern
of other biomimetic platforms with different neuron mod- recognition.
els can demonstrate the importance of the spike variation in Particularly in the work presented by Masquelier et al. [38],
neuromorphic device. it has been shown that repetition of arbitrary spatiotemporal
spike patterns hidden in spike trains can be robustly detected
and learned by multiple neurons equipped with spike-timing-
C. Digital Silicon Neuron on FPGA dependent plasticity (STDP) listening to the incoming spike
To verify the algorithms before analog implementation, we trains (Fig. 4). The neurons become selective to successive
designed a digital spiking silicon neuron (DSSN) model that coincidences of the patterns.
can be implemented in FPGA. The DSSN model is a qualita- As used in the work of Masquelier et al. [38], we generated
tive neuronal model designed for efficient implementation in spikes independently using a Poisson process with a variable
a digital arithmetic circuit and can simulate several classes of instantaneous firing rate that varies randomly between 0 and
neuronal activities: RS, FS, Low Threshold Spiking (LTS), 90 Hz. The maximal rate change was chosen so that the neuron
Intrinsically Bursting (IB), EB and SWB classes [37]. The could go from 0 to 90 Hz in 50 ms. Finally, a part of the spike
DSSN model was implemented in a Kintex-7 FPGA with trains, defined as the ‘pattern’ to be repeated, was replaced for
few resources (386 LUT and 121 FF for one FS neuron, half input lines into sections of 50ms. We randomly picked
and 401 LUT and 123 FF for RS neuron, no DSP are used). one of these sections and copy the corresponding spikes. As
Fig. 3 shows hardware implementation of FS and RS neuron. shown in Fig. 5, four different hidden patterns were gener-
Our digital system can simulate different classes of neuron ated (blue, pink, cyan and yellow) with repetition at random
in real-time. The computational cost of the DSSN model is intervals within stochastic Poissonian activity.
larger than that of the recent sophisticated Integrate-and-Fire- The criteria for recognition are the hit rate above 90% and
based models such as IZH and exponential I&F models, but the false alarm rate below 1 Hz. Using 9 DSSN neurons (LTS
smaller than ionic-conductance models. This model is intended neuron model), 4 patterns and 512 afferent neurons, the suc-
to provide a good trade-off that satisfies the demand for large- cess rate reaches 88%. Using different family of neurons (FS,
scale neuronal network simulation with biomimetic models. RS or LTS) resulted difference in hit rate (69% for RS for
580 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 65, NO. 5, MAY 2018

Fig. 6. Schematic of Scolopidia and ionic channels.

Fig. 5. Example of 4 patterns detection using 512 afferent neurons and

9 DSSN neurons. On the left side, there is raster plot of the 512 afferent
neurons. On the right side, DSSN neurons are spiking when they detect one
pattern.

512 afferent neurons and 9 output neurons) and demonstrated

that the choice of neuron family is important for such signal
processing. Fig. 7. Simulation results of the first model. (a) corresponds to GMET
This brief is expected to be used for pattern recog- variation, (b) EX shows variation of ion concentration, (c) and (d) describes
nition directly from the biological neuron recording from membrane potential VnA and VnB of neuron A and neuron B, respectively.
All variables do not have units as it is from a qualitative model.
MEA (Micro-Electrode Array) for biomedical applica-
tions [18] and for neurobiohybrid robotic applications [39].
Using this real-time pattern recognition on neuron cul-
ture recordings from MEA with healthy neurons and with channels (assumed ionic channels in Fig. 6). It is not elu-
neurological disorder neurons may help neuroscientists to cidated which types of ionic channels are expressed in the
distinguish the different behaviors on specific disease. As cilium. When the mechanosensitive ionic channels are open,
these patterns are hidden in a noisy environment (sponta- the ionic current flows into the cilium and depolarizes it, then
neous activity), this algorithm will be very efficient for this the neuron is also depolarized because the neuron and the cil-
case study. ium are electrically connected. We assumed that the mutual
inhibition of two neurons is realized by the change of ionic
composition in the scolopale space, and considered possible
B. Neural Network Mimicking the Hearing System combinations of ionic channels that can change the ionic com-
of Drosophila position in the scolopale space depending on the activation of
The nervous system of Drosophila melanogaster is com- the neurons.
posed of a relatively small number of neurons, about a hundred Our simplest model incorporates a single type of ion and
thousand neurons, whose connectivity and physiology are a voltage-gated ionic channel. The ionic composition in the
being addressed by many researchers in experimental neuro- scolopale space is not elucidated, thus we call it ion X. The
science. In spite of its simplicity, some experimental studies ion X is assumed to be filled in the scolopale space at high
revealed its intelligent behavior, for example in classical concentration and in the cilia and neurons at low concentration.
conditioning [40] and in courtship behavior using pulse sound As shown in the simulation results in Fig. 7, this model could
produced by wingbeat [41]. Johnston’s organ is a sensory produce selectivity between oscillatory and deflective stimuli.
organ in the second segment of the antennae, which detects Here GMET is a gating variable of the mechanosensitive ionic
sound, gravity, and wind [42]. It contains about two hundred channel and EX is a reversal potential of the voltage-gated
scolopidia, each of is composed of the scolopale cell, the lig- ionic channel. VnA and VnB are membrane potentials of neuron
ament cell, and two neurons (Fig. 6). Axons of the neurons A and B. While t is from 0s to 2s, an oscillatory stimulus is
project to the antennal mechanosensory and motor centers. applied and neuron A showed resonant spikes and neuron B
One of the neurons selectively responds to oscillatory stim- did not fire. The deflection is added while t is from 2s to
ulation (sound) and the other neuron responds to deflection 3s, and neuron B fired repetitively and reversal potential EX
(wind and gravity). And, it is thought that there is mutual inhi- decreased by a massive reduction of the concentration of ion
bition between the two neurons, because they do not fire at the X in the scolopale space. As the ionic current flowed into the
same time. In collaboration with Kamikouchi laboratory [42], cilium, neuron A is inhibited.
we proposed three possible models of Johnston’s organ that In the two other models, we investigated the ionic concen-
describe the selective response of neurons and the mutual tration variation in scolopale space and the one-way inhibition
inhibition. from neuron B to neuron A. In our future works, we will con-
They are composed of two pairs of a neuron and a cilium duct biological experiments on the basis of our model, and
model. We assumed that one of the neurons responding to improve models toward elucidating information processing of
oscillatory input has the electrical tuning in analogy with hair Johnston’s organ. We will implement our model first in FPGA
cells found in several species [43]. The other neuron respond- and then in analog chip. We expect to design an innovative
ing to deflection is a general spiking neuron. Each of the sensor system that responds to different kind of input with low
two cilia has mechanosensitive ionic channels and other ionic latency and low power consumption.
LEVI et al.: DEVELOPMENT AND APPLICATIONS OF BIOMIMETIC NEURONAL NETWORKS TOWARDS BMAI 581

IV. C ONCLUSION [17] S. Vassanelli and M. Mahmud, “Trends and challenges in neuro-
engineering: Toward ‘intelligent’ neuroprostheses through brain-‘brain
The main features of this brief can be summarized in five inspired systems’ communication,” Front. Neurosci., vol. 10, p. 438,
points. 1) The creation of a new neuronal model is a good Sep. 2016.
trade-off between biological plausibility and low resource [18] S. Potter, A. El Hady, and E. Fetz, “Closed-loop neuroscience and neuro-
implementation. Compared to other I&F-based models, this engineering,” Front. Neural Circuits, vol. 8, pp. 2013–2015, Sep. 2014.
[19] T. Levi, P. Bonifazi, P. Massobrio, and M. Chiappalone, “Closed-loop
model can replicate several classes of neuronal activities and systems for next-generation neuroprostheses,” Front. Neurosci., vol. 12,
also action potential shape which is important in brain sig- p. 26, Feb. 2018.
nal processing. 2) Then, we designed silicon neurons in an [20] F. D. Broccard, S. Joshi, J. Wang, and G. Cauwenberghs, “Neuromorphic
neural interfaces: From neurophysiological inspiration to biohybrid cou-
analog chip in TSMC 0.25 µm technology. This ASIC pro- pling with nervous systems,” J. Neural Eng., vol. 14, no. 4, 2017,
vides very low power consumption (5 nW) and can replicate Art. no. 041002.
many neuronal activities such as elliptic bursting mode. 3) We [21] P. Bonifazi et al., “In vitro large-scale experimental and theoretical stud-
also implemented silicon neurons in FPGA. We validated ies for the realization of bi-directional brain-prostheses,” Front. Neural
Circuits, vol. 7, p. 40, Mar. 2013.
this implementation by mimicking cortical neurons. Then we [22] S. Joucla et al., “Generation of locomotor-like activity in the iso-
used these two systems for different applications. 4) First in lated rat spinal cord using intraspinal electrical microstimulation driven
AI with a new algorithm of pattern recognition using just by a digital neuromorphic CPG,” Front. Neurosci., vol. 10, p. 67,
9 DSSN neurons with STDP and lateral inhibitory connec- Feb. 2016.
[23] A. Chiolerio, M. Chiappalone, P. Ariano, and S. Bocchini, “Coupling
tions for detecting four different patterns. This system can resistive switching devices with neurons: State of the art and perspec-
be used to characterize biological culture on MEA or in asyn- tives,” Front. Neurosci., vol. 11, p. 70, Feb. 2017.
chronous sensors. 5) Finally, we developed models for hearing [24] A. Serb et al., “A geographically distributed bio-hybrid neural network
system of Drosophila, to further understand the biological with memristive plasticity,” arXiv:1709.04179, 2017.
[25] M. Pospischil et al., “Minimal Hodgkin–Huxley type models for dif-
intelligence and to design new generation of low-power and ferent classes of cortical and thalamic neurons,” Biol. Cybern., vol. 99,
low-resource sensors. All these systems are included in the nos. 4–5, pp. 427–441, 2008.
BMAI project in collaboration with The University of Tokyo [26] D. Debanne, A. Bialowas, and S. Rama, “What are the mechanisms
for analogue and digital signalling in the brain?” Nat. Rev. Neurosci.,
and NEC Corporation. vol. 14, pp. 63–69, Nov. 2013.
[27] R. Brette, “Philosophy of the spike: Rate-based vs. spike-based theories
R EFERENCES of the brain,” Front. Syst. Neurosci., vol. 9, p. 151, Nov. 2015.
[1] G. Indiveri et al., “Neuromorphic silicon neuron circuits,” Front. [28] E. M. Izhikevich, “Simple model of spiking neurons,” IEEE Trans.
Neurosci., vol. 5, p. 73, May 2011. Neural Netw., vol. 14, no. 6, pp. 1569–1572, Nov. 2003.
[2] T. Levi and T. Fujii, “Microfluidic neurons: A new way in neuromorphic [29] W. Gerstner and W. Kistler, Spiking Neuron Models: Single Neurons,
engineering?” Micromachines, vol. 7, p. 146, Aug. 2016. Populations, Plasticity. Cambridge, U.K.: Cambridge Univ. Press, 2002.
[3] T. Levi et al., “Neuromimetic integrated circuits,” in VLSI Circuits for [30] H. Alle and J. R. P. Geiger, “Combined analog and action poten-
Biomedical Applications. London, U.K.: Artech House, 2008, ch. 12, tial coding in hippocampal mossy fibers,” Science, vol. 311, no. 5765,
pp. 241–264. pp. 1290–1293, Mar. 2006.
[4] J. Misra and I. Saha, “Artificial neural networks in hardware: A sur- [31] T. Kohno and K. Aihara, “Mathematical-model-based design method
vey of two decades of progress,” Neurocomputing, vol. 74, nos. 1–3, of silicon burst neurons,” Neurocomputing, vol. 71, nos. 7–9,
pp. 239–255, 2010. pp. 1619–1628, Mar. 2008.
[5] M. Mahowald and R. Douglas, “A silicon neuron,” Nature, vol. 354, [32] T. Kohno, M. Sekikawa, J. Li, T. Nanami, and K. Aihara, “Qualitative-
pp. 515–518, Dec. 1991. modeling-based silicon neurons and their networks,” Front. Neurosci.,
[6] R. Jung, E. J. Brauer, and J. J. Abbas, “Real-time interaction between a vol. 10, no. 273, pp. 1–16, Jun. 2016.
neuromorphic electronic circuit and the spinal cord,” IEEE Trans. Neural [33] T. Kohno, M. Sekikawa, and K. Aihara, “A configurable qualitative-
Syst. Rehabil. Eng., vol. 9, no. 3, pp. 319–326, Sep. 2001. modeling-based silicon neuron circuit,” Nonlin. Theory Appl., vol. 8,
[7] G. Le Masson, S. R.-L. Masson, D. Debay, and T. Bal, “Feedback inhibi- no. 1, pp. 25–37, Jan. 2017.
tion controls spike transfer in hybrid thalamic circuits,” Nature, vol. 417, [34] A. L. Hodgkin and A. F. Huxley, “A quantitative description of mem-
pp. 854–858, Jun. 2002. brane current and its applications to conduction and excitation in nerve,”
[8] P. Hasler, S. Kozoil, E. Farquhar, and A. Basu, “Transistor channel den- J. Physiol., vol. 117, no. 4, pp. 500–544, 1952.
drites implementing HMM classifiers,” in Proc. Int. Symp. Circuits Syst., [35] S. Brink, S. Nease, and P. Hasler, “Computing with networks of
New Orleans, LA, USA, 2007, pp. 3359–3362. spiking neurons on a biophysically motivated floating-gate based
[9] F. Grassia et al., “Tunable neuromimetic integrated system for emulating neuromorphic integrated circuit,” Neural Netw., vol. 45, pp. 39–49,
cortical neuron models,” Front. Neurosci., vol. 5, p. 134, Dec. 2011. Sep. 2013.
[10] J. Schemmel, L. Kriener, P. Muller, and K. Meier, “An accelerated analog [36] V. Rangan et al., “A subthreshold aVLSI implementation of the
neuromorphic hardware system emulating NMDA- and calcium-based Izhikevich simple neuron model,” in Proc. IEEE Eng. Med. Biol. Conf.,
non-linear dendrites,” in Proc. Int. Joint Conf. Neural Netw., Anchorage, Buenos Aires, Argentina, Aug. 2010, pp. 4164–4167.
AK, USA, May 2017, pp. 2217–2226. [37] T. Nanami and T. Kohno, “Simple cortical and thalamic neuron models
[11] N. Qiao et al., “A reconfigurable on-line learning spiking neuromorphic for digital arithmetic circuit implementation,” Front. Neurosci., vol. 10,
processor comprising 256 neurons and 128K synapses,” Front. Neurosci., no. 181, pp. 1–12, May 2016.
vol. 9, p. 141, Apr. 2015. [38] T. Masquelier, R. Guyonneau, and S. J. Thorpe, “Competitive STDP-
[12] R. Wang et al., “An FPGA implementation of a polychronous spiking based spike pattern learning,” Neural Comput., vol. 21, no. 5,
neural network with delay adaptation,” Front. Neurosci., vol. 7, p. 14, pp. 1259–1276, 2009.
Feb. 2013. [39] T. Levi et al., “Biomimetic neural network for modifying biological
[13] P. Merolla et al., “A million spiking-neuron integrated circuit with a scal- dynamics during hybrid experiments,” J. Artif. Life Robot., vol. 22, no. 3,
able communication network and interface,” Science, vol. 345, no. 6197, pp. 398–403, May 2017.
pp. 668–673, 2014. [40] T. Tully and W. G. Quinn, “Classical conditioning and retention in
[14] S. B. Furber et al., “Overview of the SpiNNaker system architecture,” normal and mutant Drosophila melanogaster,” J. Compar. Physiol. A,
IEEE Trans. Comput., vol. 62, no. 12, pp. 2454–2467, Dec. 2013. vol. 157, no. 2, pp. 263–277, Sep. 1985.
[15] A. D. Rast et al., “A location-independent direct link neuromorphic [41] A. W. Ewing and H. C. Bennet-Clark, “The courtship songs of
interface,” in Proc. Int. Joint Conf. Neural Netw., Dallas, TX, USA, Drosophila,” Behaviour, vol. 31, nos. 3–4, pp. 288–301, 1968.
2013, pp. 1–8. [42] A. Kamikouchi et al., “The neural basis of Drosophila gravity-sensing
[16] M. Ambroise, T. Levi, S. Joucla, B. Yvert, and S. Saïghi, “Real- and hearing,” Nature, vol. 458, pp. 165–171, Mar. 2009.
time biomimetic central pattern generators in an FPGA for hybrid [43] R. Fettiplace and P. A. Fuchs, “Mechanisms of hair cell tuning,” Annu.
experiments,” Front. Neurosci., vol. 7, p. 215, Nov. 2013. Rev. Physiol., vol. 61, pp. 809–834, Mar. 1999.