Professional Documents
Culture Documents
A Review On Methods, Issues and Challenges in Neuromorphic Engineering
A Review On Methods, Issues and Challenges in Neuromorphic Engineering
Mohammed Riyaz Ahmed, Member IEEE, and B.K. Sujatha, Member, IEEE
0899
This full-text paper was peer-reviewed and accepted to be presented at the IEEE ICCSP 2015 conference.
domain. Most of the processing such as filtering, spectrum 5 Transmission Synaptic activity Wires
analysis and signal compression is done before sending to
6 Hardware Brain Processor
brain. The inner ear uses just 14 microwatts and could run for
15 years [8]. The challenge is to meet this lowest power 7 Software Wetware/ Mind Operating system
consumption.
8 Memory hippocampus Gate capacitance
Despite the fact that, brain doesn't execute coded
instructions, instead uses spikes for communication between 9 Storage Monolithic Modular
synapses, collectively brain seems to be faster. A staggering 10 Logic Fuzzy Digital 0/1
amount of 10 quadrillion (1016) synaptic activity per second.
11 Connectivity High Poorly connected
The challenge is to achieve such a high connectivity and high
fan-out capability. 12 Fan-out High Very low
Human brain is fault tolerant, and is only vulnerable during Individually slow Individually fast
13 Speed
accidents and aging. It is said that von Neumann architecture Collectively fast collectively slow
Power
mimics left brain, and we are in pursuit of Right part of brain 14 Low High
consumption
which is responsible for being fault tolerant, reconfigurable
15 Reliabilty Redundant Fault-sensitive
and event driven. Learning is exhibited by this part of the
brain [9-11]. A comparison of brain's equivalent counterpart 16 Cognition Exhibited Yet to exhibit
of computers are shown in Table I.
0900
This full-text paper was peer-reviewed and accepted to be presented at the IEEE ICCSP 2015 conference.
III. IMPLEMENTATlON LEVELS OF BRAIN MODELS which it may be expensive, and the models become too
complex. In Bottom-up approach as shown in Fig. 1, we
Modelling is a reductionist approach. Models reduce
generalize one model of neuron and climb up to behavior level
complexity and provide simplified representation of real
where most of the time the models fail to replicate the
systems. All neuronal dynamics is not yet understood, hence
biological counterpart. This ambiguity whether to choose top
creating an exact replica is not under our grasp. Abstraction is
down approach (complex biological implementation, Biology
done at every level to make our life simple.
has the upper hand) or Bottom-up approach (abstract level
Brain being the most complex system, its implementation
implementations, existing engineering technology has the
can be studied in a hierarchy in Table II. The interactions of
upper hand) leads us to the Valley of Death.
different neuronal components give rise to behaviour. Since
behaviour is a collective function of different neural
TABLE IT
components, its abstraction is placed above system level.
LEVELS OF IMPLEMENTATION
Protein/Genetic level describes the genes structure. It's a
neglected/not yet explored field in Neuromorphic Engineering. Neuromorphic Correspondence for Implementation Hierarchy
At Membrane Level, electrical and ion exchange happens. levels
Hierarchy Neuroscience Electrical science
Transistor is often abstracted as switch: ON or OFF. But the
7 Behavior Level Mind Architecture
V-I characteristics shows that current flowing is smooth and
steep function of applied Voltage. Transistors work in sub 6 System Level Brain system Macro Block
threshold region where the V-I characteristics resemble the Local Neuronal
5 Circuit Level Block/Cell
current voltage relationships in molecular structures on surface population
Component
of brain cells. Hence membrane level abstractions and 4 Single Neuron Perceptron
Level
implementations are done at transistor level. CMOS/
3 Device Level Synapses
Intra-neuron communication happens through synapses. Memristors
Synaptic plasticity is the reason for the emergent animal Membrrane
2 Channel Ions Transistor
Level
behaviors: adaptation and Learning. In recent years the
learning process has dragged more attention. Selective Protein! Genetic Genes
0 t:
assigns a fixed address to every neuron, by using which S'
0
(Jq 'B
.,
neurons continuously update their central system about their 0.- Membrane Level
0 tl
excitation levels. This updated info is sent to upper/higher � �
oJ)
layers. AER is a communication protocol for spiking neurons t:
Genetic Level :Q
between different layers [15-18].This field has attracted a huge .,
f-;
community of researchers who are engineering various
protocols for inter and intra chip communication. AER scheme Fig, 1. Implementation Approaches
0901
This full-text paper was peer-reviewed and accepted to be presented at the IEEE ICCSP 2015 conference.
implement simple tasks performed by biological systems. It is possible to simulate these circuits in software. [31-42].
Neuromorphic engineering applies computational cognitive The high computational load can be taken care by parallel
Neuroscience principles discovered in living organisms to processing, but the real time environment needed for
implement such tasks in Machines using CMOS VLSI experimentation is not in the grasp of simulation.
Technology. FPGA has been popular in industry, education and research.
What's cognitive about Neuromorphic engineering? When The reasons for their overwhelming popularity is the
do we say neuromorphic systems are cognitive? What are the flexibility for fast prototyping for implementation of digital
different cognitive tasks of our interest? Neuromorphic designs without altering the hardware. Chip design demands
engineering bridges the gap between existing engineering and large waiting period and puts a hole in our pocket. FPGA is
Cognitive Neuroscience. Following are the humble attempts of continuously evolving and are available for a reasonable price.
it towards filling the gap including Neurolinguistics, vision Xilinx is one of the manufacturer. Achronix semiconductors
and auditory systems, low power devices, locomotion, are providing new class of FPGAs with speed up to 1.5 GHz.
implantable chips(biomedical), which exhibit cognitive The key change is the internal logic is asynchronous [43]. A
behaviour such as information processing, pattern direct transition from Simulink to FPGA circuit synthesis is
classification, learning and memory, decision making etc . . . possible as in [44]. To exploit the analog computation i.e.
B. Achievements ofNeuromorphic Engineering Simulink to Analog hardware conversion is possible, now the
1) Vision system: This is the most explored cognitive tasks arrays are FPAA [45, 46].
which has much interested audience not only from More often Neuromorphic engineers rely on garden-variety
medicine but also robotics and automobile field. This huge VLSI, CMOS technology to experiment as in [47] there is
vested interest is because they have immediate much resemblance between VLSI hardware and neural
applications in sensors and sensory systems. The earliest wetware.
vision systems was proposed by Mahowald and Mead
[19]. Neuromorphic engineers have made significant
VI. CONCLUSION
efforts in copying the functionality of retina by designing
analog silicon circuits which has led to wearable and It seems that we have taken a big leap, but much remains to
implantable chips [2, 3, 6, 20-24]. be achieved. Any Engineer who want to jump into research of
2) Auditory system: In 1997 researchers at Loughborough Brain Computer Interaction, Intelligent machines,
University designed a VHDL-based pitch detection Neuromorphic engineering has to know many trade-offs. This
system, and many implementations of bionic ears have paper has done a brief review of various tradeoffs to be
been carried out in [5, 6, 25]. considered during Hardware implementation of Bio-inspired
3) Microelectronic Nose: Mice and Bees demonstrate Computing Architectures. Levels of Implementation is an
powerful and efficient odor discrimination capabilities. In important contribution of this paper. Implementation
[26], a microcontroller based electronic nose is explained. approaches provide a clear idea for a researcher to start at
This is implemented using time domain encoding schemes what level and what to expect from the upper layer and what
to achieve gas identification. This pioneers translation of
services are necessary to the lower layer. This paper brings out
neurophysiological phenomena such as simple recognition,
many open research issues pertaining to morphed circuits.
easy calibration, and training into hardware. The power
Some milestones achieved are mentioned, and the things yet to
consumption by sensor is of concern.
be achieved are discussed. This has also discussed the gap
4) Brain: In [27] [28] the most ambitious project of "Human
between the software simulation results and Hardware
Brain Project" funded by European countries is aiming
emulation results.
towards many aspects of Neuromorphic Engineering such
Future work would be focusing on Learning and memory
as: Neuroinformatics, Brain simulation, High speed
computing, medical informatics, neuromorphic computing issues in implementation on Neuromorphic VLSI Chips.
and neurorobotics. SYNAPSE of DARPA [29, 30] aims to Reinforcement learning would be the focused field in
rebuild brain. Learning.
e. Open Research Issues inNeuromorphic Engineering
The modern day simulations are slow, we not only have to ACKNOWLEDGMENT
implement behaviour also need to control them. Intelligent The authors would like to thank MSRIT, REVA University
behaviour is about adaptation to any given circumstances with and JAIN University for providing necessary support and
maximum rewards in long term. Humans are active receivers Infrastructure to carry out research work.
of information i.e. perception is not always data driven, it is
Knowledge driven too. REFERENCES
0902
This full-text paper was peer-reviewed and accepted to be presented at the IEEE ICCSP 2015 conference.
[5] S. Jones, R. Meddis, S. C. Lim, and A. R. Temple, "Toward a digital [29] G. Ling, "Newsmaker interview: Geoffrey Ling. DARPA aims to
neuromorphic pitch extraction system," iEEE Trans Neural Netw, vol. rebuild brains. Interview by Emily Underwood," Science, vol. 342, pp.
II,pp. 978-87,2000. 1029-30,Nov 29 2013.
[6] S. C. Liu and T. Delbruck, "Neuromorphic sensory systems," Curr Opin [30] N. V. Thakor, D. F. Moore, R. A. Miranda, and G. S. Ling, "Special
Neurobiol, vol. 20,pp. 288-95,Jun 2010. issue of DARPA NEST proceedings," IEEE Trans Neural Syst Rehabil
[7] C. Mead, "Neuromorphic electronic systems," Proceedings of the iEEE, Eng, vol. 20,pp. 113-6,Mar 2012.
vol. 78,pp. 1629-1636, 1990. [31] D. Pecevski, D. Kappel, and Z. Jonke, "NEVESIM: event-driven neural
[8] R. Sarpeshkar, "Brain power-borrowing from biology makes for low simulation framework with a Python interface," Front Neuroinform, vol.
power computing [bionic ear]," Spectrum, iEEE, vol. 43, pp. 24-29, 8,p. 70,2014.
2006. [32] O. Gurcan, K. S. Turker, J. P. Mano, C. Bernon, O. Dikenelli, and P.
[9] A. M. Galaburda,M. LeMay,T. L. Kemper,and N. Geschwind, "Right Glize, "Mimicking human neuronal pathways in silico: an emergent
left asymmetrics in the brain," Science, vol. 199, pp. 852-856,1978. model on the effective connectivity," J Comput Neurosci, vol. 36, pp.
[10] N. A. Badian, "Nonverbal disorders of learning: The reverse of 235-57,Apr 2014.
dyslexia?," Ann Dyslexia, vol. 36,pp. 253-69,Jan 1986. [33] O. E. Hernandez and E. E. Zurek, "Teaching and learning the Hodgkin
[11] T. Okusa, R. Kakigi, and N. Osaka, "Cortical activity related to cue Huxley model based on software developed in NEURON's
invariant shape perception in humans," Neuroscience, vol. 98, pp. 615- programming language hoc," BMC Med Educ, vol. 13,p. 70,2013.
24,2000. [34] E. Arguello, R. Silva, C. Castillo, and M. Huerta, "The Neuroid: A
[12] A. L. Hodgkin and A. F. Huxley, "A quantitative description of novel and simplified neuron-model," Conf Proc iEEE Eng Med Bioi
membrane current and its application to conduction and excitation in Soc, vol. 2012,pp. 1234-7,2012.
nerve," J Physiol, vol. 117,pp. 500-44, Aug 1952. [35] T. Yamazaki,H. Ikeno, Y. Okumura,S. Satoh,Y. Kamiyama,Y. Hirata,
[13] c. Morris and H. Lecar, "Voltage oscillations in the barnacle giant et aI., "Reprint of: Simulation Platform: a cloud-based online simulation
muscle fiber," BiophysJ, vol. 35,pp. 193-213,Jul 1981. environment," Neural Netw, vol. 24,pp. 927-32,Nov 2011.
[14] E. M. lzhikevich, "Simple model of spiking neurons," iEEE Trans [36] M. L. Hines, A. P. Davison, and E. Muller, "NEURON and Python,"
Neural Netw, vol. 14,pp. 1569-72,2003. Front Neuroinform, vol. 3,p. 1,2009.
[15] R. T. Davis and J. A. Ferraro, "Comparison between AER and behavioral [37] A. P. Davison, D. Bruderle, J. Eppler, 1. Kremkow, E. Muller, D.
thresholds in normally and abnormally hearing chinchillas," Ear Hear, Pecevski, et aI., "PyNN: A Common Interface for Neuronal Network
vol. 5,pp. 153-9,May-Jun 1984. Simulators," Front Neuroinform, vol. 2,p. 11,2008.
[16] A. Linares-Barranco, G. Jimenez-Moreno, B. Linares-Barranco, and A. [38] D. Goodman and R. Brette, "Brian: a simulator for spiking neural
Civit-Balcells, "On algorithmic rate-coded AER generation," IEEE networks in python," Front Neuroinform, vol. 2,p. 5,2008.
Trans Neural Netw, vol. 17,pp. 771-88,May 2006. [39] C. Moulin, A. Gliere, D. Barbier, S. Joucla, B. Yvert, P. Mailley, et aI.,
[17] C. Zamarreno-Ramos, T. Serrano-Gotarredona, and B. Linares "A new 3-D finite-element model based on thin-film approximation for
Barranco, "A 0.35 mum sub-ns wake-up time ON-OFF switchable microelectrode array recording of extracellular action potential," 1EEE
LVDS driver-receiver chip VO pad pair for rate-dependent power saving Trans Biomed Eng, vol. 55,pp. 683-92,Feb 2008.
in AER bit-serial links," IEEE Trans Biomed Circuits Syst, vol. 6, pp. [40] Y. Hayashida, "[Computer simulation using the "Neuron ": a kinetic
486-97,Oct 2012. model for channels based on the Markov state transition process],"
[18] B. Zhao, R. Ding, S. Chen, B. Linares-Barranco, and H. Tang, Nihon Seirigaku Zasshi, vol. 68,pp. 319-24,2006.
"Feedforward Categorization on AER Motion Events Using Cortex-Like [41] J. F. Vibert, K. Pakdaman, E. Boussard, and E. Av-Ron, "XNBC: a
Features in a Spiking Neural Network," IEEE Trans Neural Netw Learn simulation tool. Application to the study of neural coding using hybrid
Syst, Oct 22 2014. networks," Biosystems, vol. 40,pp. 211-8,1997.
[19] M. A. Mahowald and C. Mead, "The silicon retina," Sci Am, vol. 264, [42] B. Ans, 1. C. Gilhodes, and J. Herault, "[Simulation of neuronal
pp. 76-82,May 1991. networks (SIRENE). 11. Hypothesis for decoding the message of
[20] K. A. Zaghloul and K. Boahen, "A silicon retina that reproduces signals movement carried by spindle afferences IA and II by a mechanism of
in the optic nerve,"J Neural Eng, vol. 3,pp. 257-67,Dec 2006. synaptic plasticity]," C R Seances Acad Sci iii, vol. 297, pp. 419-22,
[21] T. Delbruck, A. van Schaik, and 1. Hasler, "Research topic: 1983.
neuromorphic engineering systems and applications. A snapshot of [43] R. Paz-Vicente, E. Cerezuela-Escudero, M. Dominguez-Morales, A.
neuromorphic systems engineering," Front Neurosci, vol. 8, p. 424, Jimenez-Fernandez, A. Linares-Barranco, and G. Jimenez-Moreno, "A
2014. perfomance comparison study between synchronous and asynchronous
[22] G. Indiveri, B. Linares-Barranco, T. J. Hamilton, A. van Schaik, R. FPGA for spike based systems. Under the AER synthetic generation," in
Etienne-Cummings, T. Delbruck, et aI., "Neuromorphic silicon neuron Performance Evaluation of Computer & Telecommunication Systems
circuits," Front Neurosci, vol. 5,p. 73,2011. (SPECTS), 2011 international Symposium on, 2011,pp. 38-45.
[23] R. Serrano-Gotarredona,M. Oster,P. Lichtsteiner,A. Linares-Barranco, [44] B. Sbarcea and D. Nicula, "Automatic conversion of matlab/simulink
R. Paz-Vicente, F. Gomez-Rodriguez, et aI., "CAVIAR: a 45k neuron, models to hdl models," in international Conference on Optimization of
5M synapse, 12G connects/s AER hardware sensory-processing Electrical and Electronic Equipment, 2004,pp. 67-70.
learning-actuating system for high-speed visual object recognition and [45] F. Baskaya, S. Reddy, S. K. Lim, and D. V. Anderson, "Placement for
tracking," iEEE Trans Neural Netw, vol. 20,pp. 1417-38,Sep 2009. large-scale floating-gate field-programable analog arrays," Very Large
[24] F. Zhengming, T. Delbruck, P. Lichtsteiner, and E. Culurciello, "An Scale integration (VLSi) Systems, IEEE Transactions on, vol. 14, pp.
address-event fall detector for assisted living applications," IEEE Trans 906-910,2006.
Biomed Circuits Syst, vol. 2,pp. 88-96,Jun 2008. [46] F. Baskaya, B. Gestner, C. Twigg, S. K. Lim, D. V. Anderson, and P.
[25] P. K. Park, H. Ryu, J. H. Lee, C. W. Shin, K. B. Lee, J. Woo, et aI., Hasler, "Rapid prototyping of large-scale analog circuits with field
"Fast neuromorphic sound localization for binaural hearing aids," Conf programmable analog array," in Field-Programmable Custom
Proc IEEE Eng Med Bioi Soc, vol. 2013,pp. 5275-8,2013. Computing Machines, 2007. FCCM 2007. 15th Annual IEEE
[26] C. Hung Tat, N. Kwan Ting, A. Bermak, M. K. Law, and D. Martinez, Symposium on, 2007,pp. 319-320.
"Spike latency coding in biologically inspired microelectronic nose," [47] L. Jihong and W. Chengyuan, "A Survey of Neuromorphic Engineering
iEEE Trans Biomed Circuits Syst, vol. 5,pp. 160-8,Apr 20II. -Biological Nervous Systems Realized on Silicon," in Testing and
[27] A. Calimera, E. Macii, and M. Poncino, "The Human Brain Project and Diagnosis, 2009. ICTD 2009. iEEE Circuits and Systems international
neuromorphic computing," Funct Neurol, vol. 28, pp. 191-6, Jul-Sep Conference on, 2009,pp. 1-4.
2013.
[28] A. Schwartz, "First volley in the brain race? Europe's human brain
project the first to the starting line--will U.S. brain initiative catch up?,"
Ann Neurol, vol. 73,p. A7,Jun 2013.
0903