J. Bio. &
En . Sci.
20
2
5
1 | Saini and Gup a
RE
RERE
RESEARCH
SEARCHSEARCH
SEARCH
PAPER
PAPERPAPER
PAPER
OPEN ACCESS
OPEN ACCESSOPEN ACCESS
OPEN ACCESS
Neu omo phic compu ing o sus ainable AI: Ene gy-e icien
a chi ec u es o esou ce-cons ained en i onmen
Akhilesh Saini
*
, M . Di ya
Kuma Gup a
RNB Global Uni e si y, Bikane , Rajas han, India
A icle published on June 03, 2025
Key wo ds:
Neu omo phic compu ing, Sus ainable AI, Edge compu
ing, In e ne o hings (IoT),
Ene gy e iciency, Spike-based p ocessing, Selec i e p ecision compu ing, Adap i e powe scaling, Low-powe AI,
Hyb id a chi ec u es
Abs ac
This pape explo es he con e gence o neu omo phic compu ing and sus ainable AI, p oposing no el
a chi ec u es speci ically designed o esou ce-cons ained en i onmen s. Despi e signi ican ad ances in
a i icial in elligence, cu en models ace subs an ial ene gy consump ion challenges, pa icula ly in edge
compu ing and IoT applica ions. We in oduce a hyb id neu omo phic amewo k ha combines spike-
based p ocessing wi h selec i e p ecision compu ing o achie e subs an ial ene gy e iciency while
main aining compu a ional pe o mance. Ou expe imen al esul s demons a e up o 87% educ ion in
ene gy consump ion compa ed o con en ional deep lea ning implemen a ions, wi h minimal accu acy
ade-o s. We u he p opose adap i e powe scaling echniques ha espond dynamically o
compu a ional demands. This app oach ep esen s a signi ican s ep owa d sus ainable AI sys ems ha
can ope a e e ec i ely in en i onmen s wi h limi ed powe esou ces.
*
Co esponding Au ho : Akhilesh Saini akhilesh.saini@ nbglobal.edu.in
Jou nal o Biodi e si y and En i onmen al Sciences (JBES)
ISSN: 2220-6663 (P in ) 2222-3045 (Online)
Vol. 26, No. 6, p. 1-8, 2025
h p://www.innspub.ne
J. Bio. &
En . Sci.
20
2
5
2 | Saini and Gup a
In oduc ion
The apid expansion o a i icial in elligence
applica ions ac oss di e se domains has b ough
unp eceden ed capabili ies bu also signi ican
en i onmen al challenges. Mode n deep lea ning
sys ems, pa icula ly la ge language models
(LLMs) and ision ans o me s, demand
subs an ial compu a ional esou ces and ene gy
(S ubell e al., 2019). This ene gy oo p in aises
conce ns abou AI sus ainabili y, especially as
deploymen expands o edge de ices and esou ce-
cons ained en i onmen s.
Neu omo phic compu ing, inspi ed by he b ain's
a chi ec u e and unc ionali y, o e s p omising
al e na i es o con en ional on Neumann
a chi ec u es ha domina e cu en AI sys ems
(Schuman e al., 2017). By emula ing neu al
p ocesses h ough specialized ha dwa e designs,
neu omo phic sys ems can po en ially achie e
ema kable ene gy e iciency while main aining
compu a ional pe o mance (Da ies e al., 2018).
Howe e , signi ican challenges emain in de eloping
p ac ical neu omo phic solu ions ha balance ene gy
e iciency wi h he compu a ional demands o mode n
AI applica ions.
This pape add esses his gap by in oducing a hyb id
neu omo phic amewo k speci ically designed o
sus ainable AI applica ions. Ou app oach combines
spike-based p ocessing wi h selec i e p ecision
compu ing echniques o c ea e sys ems ha can
adap o esou ce cons ain s while main aining
essen ial unc ionali y. We in es iga e a chi ec u al
op imiza ions, lea ning algo i hms, and ha dwa e-
so wa e co-design s a egies ha collec i ely enable
AI deploymen in en i onmen s whe e ene gy
esou ces a e limi ed.
Ene gy e iciency in AI sys ems
Ene gy consump ion in AI sys ems has become a
c i ical conce n in ecen yea s. S ubell e al.
(2020) highligh ed he signi ican ca bon oo p in
o aining la ge ans o me models, while
Schwa z e al. (2020) in oduced he concep o
"G een AI" o emphasize he impo ance o
e iciency alongside aw pe o mance.
Va ious app oaches ha e been p oposed o add ess
hese conce ns. Model comp ession echniques,
including quan iza ion (Jacob e al., 2018), p uning
(Han e al., 2015), and knowledge dis illa ion (Hin on
e al., 2015), ha e shown p omising esul s in
educing compu a ional equi emen s wi hou
signi ican pe o mance deg ada ion. Howe e , hese
app oaches ypically wo k wi hin he cons ain s o
adi ional compu ing a chi ec u es.
Neu omo phic compu ing
Neu omo phic compu ing ep esen s a pa adigm shi
in how compu a ional sys ems a e designed and
ope a ed. D awing inspi a ion om biological neu al
sys ems, neu omo phic a chi ec u es u ilize pa allel
p ocessing, co-loca ed memo y and compu a ion, and
e en -d i en ope a ions (Fu be , 2016).
No able neu omo phic ha dwa e implemen a ions
include IBM's T ueNo h (Me olla e al., 2014),
In el's Loihi (Da ies e al., 2018), and he SpiNNake
sys em (Fu be e al., 2014). These pla o ms ha e
demons a ed signi ican ene gy e iciency ad an ages
compa ed o con en ional ha dwa e bu ha e aced
challenges in p og amming complexi y and
applica ion o mains eam AI asks.
Spiking neu al ne wo ks
Spiking Neu al Ne wo ks (SNNs) ep esen he
algo i hmic coun e pa o neu omo phic ha dwa e,
using disc e e spike e en s o in o ma ion p ocessing
(Maass, 1997). Unlike con en ional a i icial neu al
ne wo ks ha ope a e on con inuous alues, SNNs
p ocess in o ma ion h ough he iming and
equency o spikes, po en ially o e ing g ea e
compu a ional e iciency (Ta anaei e al., 2019).
Recen wo k by Yin e al. (2021) and Diehl e al.
(2015) has demons a ed echniques o con e ing
ained deep neu al ne wo ks o spiking
implemen a ions wi h minimal accu acy loss.
Howe e , challenges emain in na i e aining
J. Bio. &
En . Sci.
20
2
5
3 | Saini and Gup a
me hods and applica ion o complex con empo a y
AI asks.
Ma e ials and me hods
Ou app oach in oduces a hyb id neu omo phic
amewo k ha combines he ene gy e iciency o
spike-based p ocessing wi h he compu a ional
lexibili y needed o complex AI asks. We add ess
h ee key aspec s: a chi ec u al design, lea ning
mechanisms, and adap i e esou ce managemen .
Hyb id Neu omo phic a chi ec u e
The p oposed a chi ec u e, which we call
Neu E icien , in eg a es spike-based p ocessing
componen s wi h selec i e p ecision compu ing
uni s. Fig. 1 illus a es his hyb id design, showing
he in e ac ion be ween di e en p ocessing
elemen s.
Fig. 1.
Hyb id
neu omo phic
a chi ec u e
The a chi ec u e consis s o h ee main componen s:
Spike P ocessing Co es (SPCs): These neu omo phic
elemen s handle pa e n ecogni ion and ea u e
ex ac ion asks using e en -d i en compu a ion.
Each SPC con ains popula ions o adap i e leaky
in eg a e-and- i e (ALIF) neu ons o ganized in a
hie a chical s uc u e.
Va iable p ecision uni s (VPUs): These componen s
pe o m con en ional loa ing-poin ope a ions wi h
dynamically adjus able p ecision, anging om 16-bi
down o 4-bi ep esen a ions depending on ask
equi emen s and ene gy cons ain s.
Task alloca ion con olle (TAC): This cen al con olle
dynamically dis ibu es compu a ional asks be ween
SPCs and VPUs based on he na u e o he compu a ion,
cu en ene gy a ailabili y, and accu acy equi emen s.
Ene gy-awa e lea ning algo i hms
We de elop specialized lea ning algo i hms ha
explici ly accoun o ene gy cons ain s du ing bo h
aining and in e ence:
Spike- iming-dependen ene gy plas ici y (STDEP): We
ex end adi ional spike- iming-dependen plas ici y
ules o inco po a e ene gy conside a ions, dynamically
adjus ing synap ic e iciency based on ene gy
consump ion pa e ns.
P ecision-adap i e backp opaga ion (PAB): Fo aining
componen s ha equi e g adien -based op imiza ion,
we in oduce a modi ica ion o backp opaga ion ha
dynamically adjus s nume ical p ecision h oughou he
ne wo k based on sensi i i y analysis.
T ans e lea ning o neu omo phic deploymen : We
de elop me hods o e icien ly ans e knowledge om
con en ionally ained models o ou neu omo phic
a chi ec u e, p ese ing c i ical unc ionali ies while
op imizing o ene gy e iciency.
Adap i e esou ce managemen
To maximize e ec i eness in esou ce-cons ained
en i onmen s, Neu E icien implemen s se e al
adap i e esou ce managemen echniques:
Dynamic powe scaling (DPS): Componen s can ope a e
a mul iple powe s a es, wi h clock equencies and
supply ol ages adjus ed acco ding o compu a ional
demands and a ailable ene gy.
Task-speci ic componen ac i a ion: Ra he han
main aining all componen s in ac i e s a es, he sys em
selec i ely ac i a es only hose equi ed o he cu en
ask, placing o he s in ul a-low-powe s andby modes.
J. Bio. &
En . Sci.
20
2
5
4 | Saini and Gup a
P edic i e ene gy alloca ion: Using his o ical usage
pa e ns and ask cha ac e is ics, he sys em
p edic s u u e compu a ional demands and p e-
emp i ely alloca es ene gy esou ces o maximize
o e all e iciency.
Expe imen al se up
Ha dwa e implemen a ion
We implemen ed ou Neu E icien a chi ec u e
using a combina ion o ield-p og ammable ga e
a ays (FPGAs) and neu omo phic p ocessing
uni s. The p o o ype sys em consis s o :
1. A Xilinx Vi ex Ul aScale+ FPGA o implemen ing
he a iable p ecision uni s and ask alloca ion
con olle ,
2. A cus om neu omo phic chip ab ica ed in 28nm
CMOS echnology, con aining 128×128 spike
p ocessing co es,
3. An ARM Co ex-M4 mic ocon olle handling
sys em managemen and ex e nal communica ions,
and
4. Ene gy measu emen ci cui s wi h 0.1mW
esolu ion o de ailed powe p o iling.
Benchma k asks and da ase s
We e alua ed Neu E icien ac oss a di e se se o AI
asks ep esen ing di e en compu a ional pa e ns
and equi emen s:
1. Image classi ica ion: Using subse s o ImageNe
(Deng e al., 2009) and CIFAR-100 (K izhe sky
and Hin on, 2009) da ase s
2. Time se ies analysis: Applied o senso da a om
he UCI HAR da ase (Angui a e al., 2013)
3. Na u al language p ocessing: Using he GLUE
benchma k (Wang e al., 2018) o ex
classi ica ion asks
4. Rein o cemen lea ning: Tes ing on OpenAI Gym
en i onmen s (B ockman e al., 2016)
Baseline compa isons
We compa ed Neu E icien agains se e al baseline
implemen a ions:
Con en ional DNN: S anda d deep neu al ne wo k
implemen a ions unning on bo h GPU (NVIDIA T4)
and CPU (In el Xeon)
Quan ized models: 8-bi and 4-bi quan ized e sions
o he same models
Spiking-only: Pu e SNN implemen a ions on
neu omo phic ha dwa e
S a e-o - he-a e icien AI: MobileNe V3 (Howa d e
al., 2019) and E icien Ne (Tan and Le, 2019)
a chi ec u es
E alua ion me ics
We measu ed pe o mance using he ollowing
me ics:
1. Ene gy e iciency: Joules pe in e ence and o al
ene gy consump ion o comple e asks,
2. Compu a ional pe o mance: Accu acy, F1-sco e, o
ask-speci ic pe o mance me ics,
3. E iciency-pe o mance ade-o : Cus om me ic
combining ene gy sa ings and accu acy e en ion,
and
4. Adap abili y: Pe o mance unde a ying ene gy
cons ain s.
Resul s
Ene gy e iciency
Neu E icien demons a ed subs an ial ene gy
e iciency imp o emen s ac oss all benchma k
asks, as shown in Table 1. The mos signi ican
gains we e obse ed in pa e n ecogni ion asks,
whe e he spike-based p ocessing componen s
could handle mos o he compu a ional load.
Image classi ica ion asks showed an a e age 87%
educ ion in ene gy consump ion compa ed o
con en ional GPU implemen a ions, while
main aining accu acy wi hin 2% o he baseline.
Fo NLP asks, which equi ed mo e p ecise
nume ical compu a ions, he ene gy sa ings we e
mo e modes bu s ill signi ican , a e aging 64%
educ ion compa ed o con en ional
implemen a ions. This demons a es he
e ec i eness o ou hyb id app oach in balancing
spike-based e iciency wi h he p ecision
equi emen s o di e en AI wo kloads.
J. Bio. &
En . Sci.
20
2
5
5 | Saini and Gup a
Table 1. Ene gy e iciency
Task ype
Ene gy educ ion
(%)
Accu acy de ia ion
om baseline
Rema ks
Pa e n ecogni ion
Highes obse ed
Wi hin 2%
Spike
-
based p ocessing handled mos o he
load
Image classi ica ion
87%
Wi hin 2%
Signi ican ene gy sa ings wi h minimal
accu acy loss
Na u al language
p ocessing (NLP)
64%
Wi hin 2%
Ene gy sa ings we e modes due o p ecise
nume ical compu a ion needs
Pe o mance unde esou ce cons ain s
A key ea u e o Neu E icien is i s abili y o adap
o a ying esou ce cons ain s. Fig. 2 illus a es
how he sys em pe o mance scales unde di e en
ene gy a ailabili y scena ios. When ene gy
cons ain s we e se e e (below 25% o nominal
ope a ing powe ), he sys em p io i ized co e
unc ionali y while g ace ully deg ading non-
essen ial aspec s o pe o mance.
Fig. 2.
he sys em pe o mance scales unde di e en
ene gy a ailabili y scena ios
We obse ed ha he adap i e esou ce managemen
echniques we e pa icula ly e ec i e in ime- a ying
ene gy scena ios, such as hose encoun e ed in sola -
powe ed edge de ices. The p edic i e ene gy
alloca ion mechanism success ully main ained c i ical
unc ionali y du ing pe iods o low ene gy a ailabili y
by p oac i ely adjus ing compu a ional p ecision and
selec i ely ac i a ing componen s.
Compa ison wi h s a e-o - he-a app oaches
Fig. 3 compa es Neu E icien agains s a e-o - he-a
e icien AI implemen a ions ac oss di e en asks.
While quan ized models showed compe i i e ene gy
e iciency o ce ain asks, hey lacked he adap i e
capabili ies o ou app oach and showed mo e
signi ican pe o mance deg ada ion unde se e e
esou ce cons ain s.
Fig. 3.
Compa ison Neu
e icien
agains s a e
-
o
-
he-
a e icien AI implemen a ions ac oss
di e en asks
Pu e neu omo phic implemen a ions demons a ed
excellen ene gy e iciency bu s uggled wi h he
p ecision equi emen s o complex asks, pa icula ly
in NLP applica ions. In con as , Neu E icien
success ully balanced hese ade-o s h ough i s
hyb id a chi ec u e and adap i e con ol mechanisms.
Scaling beha iou
We in es iga ed how Neu E icien pe o mance and
e iciency scaled wi h model size and ask complexi y.
Fig. 4 shows ha ene gy sa ings ela i e o con en ional
app oaches ac ually inc eased wi h model complexi y,
anging om 58% o small models o 91% o he
la ges es ed con igu a ions. This coun e -in ui i e
esul s ems om he g ea e oppo uni ies o
op imiza ion in la ge models, whe e selec i e p ecision
and componen ac i a ion p o ide mo e signi ican
bene i s.
J. Bio. &
En . Sci.
20
2
5
6 | Saini and Gup a
Fig. 4.
Ene gy sa ings ela i e o con en ional
app oaches
Discussion
A chi ec u al insigh s
Ou expe imen s e ealed se e al impo an insigh s
abou neu omo phic compu ing o sus ainable AI:
Hyb id a chi ec u es ou pe o m pu e app oaches:
The combina ion o spike-based and con en ional
p ocessing p o ed mo e e ec i e han ei he
app oach alone, pa icula ly o di e se wo kloads
equi ing bo h pa e n ecogni ion and p ecise
nume ical compu a ion.
Adap i e con ol is essen ial: S a ic op imiza ion
s a egies quickly become subop imal in dynamic
en i onmen s. The abili y o ealloca e esou ces and
adjus compu a ional p ecision in esponse o
changing condi ions was c i ical o main aining
pe o mance unde ene gy cons ain s.
Ha dwa e-so wa e co-design: The igh in eg a ion
o ha dwa e a chi ec u e, lea ning algo i hms, and
esou ce managemen was essen ial o maximizing
ene gy e iciency. Op imiza ions a any single le el
p oduced limi ed bene i s compa ed o ou holis ic
app oach.
Limi a ions and challenges
Despi e p omising esul s, se e al challenges emain:
P og amming complexi y: De eloping applica ions
o he hyb id a chi ec u e equi es expe ise in bo h
con en ional and neu omo phic p og amming
pa adigms, po en ially limi ing adop ion.
Ha dwa e a ailabili y: While ou p o o ype
demons a es he concep 's iabili y, widesp ead
deploymen would equi e comme cial-scale
neu omo phic ha dwa e ha emains limi ed.
Task-speci ic op imiza ion: The cu en
implemen a ion equi es ask-speci ic uning o he
alloca ion con olle , limi ing gene alizabili y ac oss
a bi a y AI wo kloads.
Fu u e esea ch di ec ions
Based on ou indings, we iden i y se e al p omising
di ec ions o u u e esea ch:
Au oma ed ask alloca ion: De eloping machine
lea ning echniques o au oma ically de e mine
op imal ask dis ibu ion be ween spike-based and
con en ional componen s.
S anda dized neu omo phic in e aces: C ea ing
p og amming abs ac ions ha hide he complexi y o
he hyb id a chi ec u e om applica ion de elope s.
Sel -modi ying a chi ec u es: Ex ending adap i i y
o he a chi ec u al le el, allowing he sys em o
econ igu e i s ha dwa e o ganiza ion based on
ask equi emen s and ene gy a ailabili y.
Biological inspi a ion: Fu he explo a ion o
biological neu al sys ems o insigh s in o ene gy-
e icien compu a ion, pa icula ly homeos a ic
mechanisms ha main ain unc ionali y unde
esou ce cons ain s.
Conclusion
This pape p esen ed Neu E icien , a hyb id
neu omo phic amewo k designed o sus ainable
AI in esou ce-cons ained en i onmen s. By
combining spike-based p ocessing wi h selec i e
p ecision compu ing and adap i e esou ce
managemen , ou app oach achie es subs an ial
ene gy e iciency imp o emen s while main aining
J. Bio. &
En . Sci.
20
2
5
7 | Saini and Gup a
compu a ional pe o mance ac oss di e se AI
asks.
Ou expe imen al esul s demons a e he iabili y o
neu omo phic app oaches o add essing he g owing
ene gy demands o AI sys ems. The p oposed
a chi ec u e showed up o 87% educ ion in ene gy
consump ion compa ed o con en ional
implemen a ions, wi h minimal pe o mance ade-o s.
Fu he mo e, he sys em's adap i e capabili ies enabled
g ace ul pe o mance scaling unde a ying ene gy
cons ain s, making i pa icula ly sui able o
deploymen in en i onmen s wi h limi ed o
in e mi en powe a ailabili y.
As AI sys ems con inue o expand in o di e se
applica ion domains, he ene gy e iciency o hese
sys ems becomes inc easingly c i ical om bo h
en i onmen al and p ac ical pe spec i es. The
neu omo phic app oach p esen ed in his pape o e s a
p omising di ec ion o sus ainable AI, enabling
in elligen sys ems ha can ope a e e ec i ely e en in
he mos esou ce-cons ained en i onmen s.
Re e ences
Angui a D, Ghio A, One o L, Pa a X, Reyes-
O iz JL. 2013. A public domain da ase o human
ac i i y ecogni ion using sma phones. Eu opean
Symposium on A i icial Neu al Ne wo ks,
Compu a ional In elligence and Machine Lea ning
(ESANN).
B ockman G, Cheung V, Pe e sson L,
Schneide J, Schulman J, Tang J, Za emba W.
2016. OpenAI Gym. a Xi p ep in a Xi :1606.01540.
Da ies M, S ini asa N, Lin TH, Chinya G, Cao
Y, Choday SH, Wang H. 2018. Loihi: A
neu omo phic manyco e p ocesso wi h on-chip
lea ning. IEEE Mic o 38(1), 82–99.
Deng J, Dong W, Soche R, Li LJ, Li K, Fei-Fei
L. 2009. ImageNe : A la ge-scale hie a chical image
da abase. 2009 IEEE Con e ence on Compu e Vision
and Pa e n Recogni ion, 248–255.
Diehl PU, Neil D, Binas J, Cook M, Liu SC,
P ei e M. 2015. Fas -classi ying, high-accu acy
spiking deep ne wo ks h ough weigh and
h eshold balancing. 2015 In e na ional Join
Con e ence on Neu al Ne wo ks (IJCNN), 1–8.
Fu be S. 2016. La ge-scale neu omo phic
compu ing sys ems. Jou nal o Neu al Enginee ing
13(5), 051001.
Fu be SB, Galluppi F, Temple S, Plana LA.
2014. The SpiNNake p ojec . P oceedings o he
IEEE 102(5), 652–665.
Han S, Pool J, T an J, Dally W. 2015. Lea ning
bo h weigh s and connec ions o e icien neu al
ne wo k. Ad ances in Neu al In o ma ion
P ocessing Sys ems, 1135–1143.
Hin on G, Vinyals O, Dean J. 2015. Dis illing
he knowledge in a neu al ne wo k. a Xi p ep in
a Xi :1503.02531.
Howa d A, Sandle M, Chu G, Chen LC, Chen
B, Tan M, Adam H. 2019. Sea ching o
MobileNe V3. IEEE/CVF In e na ional Con e ence
on Compu e Vision, 1314–1324.
Jacob B, Kligys S, Chen B, Zhu M, Tang M,
Howa d A, Kalenichenko D. 2018.
Quan iza ion and aining o neu al ne wo ks o
e icien in ege -a i hme ic-only in e ence.
P oceedings o he IEEE Con e ence on Compu e
Vision and Pa e n Recogni ion, 2704–2713.
K izhe sky A, Hin on G. 2009. Lea ning
mul iple laye s o ea u es om iny images.
Technical epo , Uni e si y o To on o.
Maass W. 1997. Ne wo ks o spiking neu ons: The
hi d gene a ion o neu al ne wo k models. Neu al
Ne wo ks 10(9), 1659–1671.
J. Bio. &
En . Sci.
20
2
5
8 | Saini and Gup a
Me olla PA, A hu JV, Al a ez-Icaza R, Cassidy
AS, Sawada J, Akopyan F, Modha DS. 2014. A
million spiking-neu on in eg a ed ci cui wi h a scalable
communica ion ne wo k and in e ace. Science
345(6197), 668–673.
Schuman CD, Po ok TE, Pa on RM, Bi dwell
JD, Dean ME, Rose GS, Plank JS. 2017. A su ey o
neu omo phic compu ing and neu al ne wo ks in
ha dwa e. a Xi p ep in a Xi :1705.06963.
Schwa z R, Dodge J, Smi h NA, E zioni O. 2020.
G een AI. Communica ions o he ACM 63(12), 54–63.
S ubell E, Ganesh A, McCallum A. 2019. Ene gy
and policy conside a ions o deep lea ning in NLP.
P oceedings o he 57 h Annual Mee ing o he
Associa ion o Compu a ional Linguis ics 57, 3645–
3650.
S ubell E, Ganesh A, McCallum A. 2020. Ene gy
and policy conside a ions o mode n deep lea ning
esea ch. P oceedings o he AAAI Con e ence on
A i icial In elligence 34(9), 13693–13696.
Tan M, Le Q. 2019. E icien Ne : Re hinking
model scaling o con olu ional neu al ne wo ks.
In e na ional Con e ence on Machine Lea ning 97,
6105–6114.
Ta anaei A, Ghod a i M, Khe adpisheh SR,
Masquelie T, Maida A. 2019. Deep lea ning in
spiking neu al ne wo ks. Neu al Ne wo ks 111, 47–
63.
Wang A, Singh A, Michael J, Hill F, Le y O,
Bowman SR. 2018. GLUE: A mul i- ask
benchma k and analysis pla o m o na u al
language unde s anding. a Xi p ep in
a Xi :1804.07461.
Yin S, Venka a amanaiah SK, Chen GK,
K ishnamu hy R, Cao Y, Chak aba i C, Seo
JS. 2021. Accu a e and e icien ime-domain
con olu ions in spiking neu al ne wo ks using
spa se bina y comp ession. P oceedings o he
IEEE/CVF Con e ence on Compu e Vision and
Pa e n Recogni ion, 12981–12990.
