A compu a ional app oach o a neu omo phic
sequen ial memo y bio-inspi ed on he
Hippocampus and En o hinal Co ex o ma ion
Daniel Casanue a-Mo a o1[0000−0002−7676−1629], Al a o
Ayuso-Ma inez1[0000−0002−0059−6647] An onio M.
P´e ez-Pe˜na1[0009−0005−9754−1659], Juan P.
Dominguez-Mo ales1[0000−0002−5474−107X], and Gab iel
Jimenez-Mo eno1[0000−0003−4512−6750]
Uni e sidad de Se illa, Se illa 41004, Espa˜na,
[email p o ec ed]
Abs ac . The b ain is conside ed one o he mos powe ul and e icien
machines in exis ence. This is why neu omo phic enginee ing is ying
o mimic biology o de elop new sys ems ha inco po a e hese supe io
capabili ies. Wi hin his ield, bio-inspi ed lea ning and memo y sys ems
a e s ill a challenge o be sol ed, and his is whe e he hippocampus and
en o hinal co ex is in ol ed. This b ain o ma ion ac s as a sho - e m
memo y capable o lea ning by sel -associa ion o memo ies om di e -
en sou ces o in o ma ion. In his wo k, we p opose a ully unc ional
bio-inspi ed spike-based hippocampus and en o hinal co ex sequen ial
memo y sys em capable o lea ning memo ies, ecalling indi idual and
sequen ial memo ies om a single agmen , and o ge ing hem. This
model has been implemen ed on SpiNNake using Spiking Neu al Ne -
wo ks, and a se o expe imen s we e pe o med o demons a e i s co -
ec ope a ion. This model will pa e he oad o he de elopmen o
u u e mo e complex neu omo phic sys ems wi h a la ge amoun o ap-
plicabili y.
Keywo ds: Hippocampus model, Sequen ial memo y, Spiking Neu al
Ne wo ks, Neu omo phic enginee ing, SpiNNake
1 In oduc ion
The e olu ion o echnology has allowed o he in eg a ion o compu ing in a
wide ange o ields and applica ions. This expansion has b ough wi h i he need
o sys ems wi h highe compu a ional and memo y capaci ies. The app oach o
a digi al solu ion leads o e y high ene gy consump ion equi emen s [1, 2]. In
iew o his, he e a e di e en ields ha add ess his p oblem. Among hem all,
neu omo phic enginee ing s ands ou [2]. This ield ocuses on he p oposal o
bio-inspi ed sys ems based on na u e and, especially, on he b ain, o inco po a e
he capaci y o sol e complex p oblems e icien ly in o cu en sys ems [3,4].
2 Daniel Casanue a-Mo a o e al.
In gene al, hese sys ems p esen an app oach based on Spiking Neu al Ne -
wo ks (SNNs). These a e ne wo ks o a i icial neu ons in e connec ed by synapses
ha communica e be ween hem by means o asynch onous e en s (also called
spikes) and lea n based on synap ic plas ici y. Due o his e en -d i en ope a-
ion, neu omo phic sys ems ha e highe eal- ime compu a ional capabili ies [5],
lowe ene gy consump ion [6] and highe noise ole ance [7] han cu en digi al
sys ems.
O all he egions o he b ain, i is necessa y o highligh he hippocampus.
The hippocampus is he egion capable o lea ning and s o ing la ge amoun s
o in o ma ion om di e en sou ces quickly. I is also capable o ecalling and
ep oducing his in o ma ion sequen ially om a agmen o he memo y [8].
This sequen ial lea ning and ecall o memo ies is he key o long- e m lea ning
in animals and is o g ea u ili y and applicabili y [4].
The hippocampus has achie ed his unc ionali y h ough i s beha iou as an
au oassocia i e memo y. The in o ma ion a i es a he Den a e Gy us (DG),
which is esponsible o inc easing he dispe sion o he con en o his in o -
ma ion o acili a e lea ning and s o age. The dispe sed in o ma ion a i es a
Co nu Ammonis 3 (CA3) whe e lea ning and s o age akes place. CA3 is a e-
cu en colla e al ne wo k s uc u e ha associa es in o ma ion wi h each o he
o o m sequences o memo ies. Finally, hese memo ies lea e he hippocampus
h ough Co nu Ammonis 1 (CA1), which pe o ms he e e se ope a ion o DG
o e ie e he o iginal o ma o he in o ma ion. In addi ion, he hippocam-
pus is able o ecall hese memo ies based on a agmen o i sel . When his
agmen o a memo y eaches CA3, he ne wo k begins o oscilla e un il he
comple e memo y is e ie ed [8].
This unc ioning would no be possible wi hou he En o hinal Co ex (EC)
which ac s as he main inpu and ou pu pa hway o he hippocampus. This
egion is in ol ed in he sequen ial ecall o memo ies by ac ing as a second
ecu en ne wo k. Speci ically, EC laye 5 (EC5) ecei es ou pu in o ma ion
om he hippocampus and o wa ds i o EC laye 2 (EC2). EC2, in u n, sends
some o his in o ma ion back o he hippocampus, s a ing he ecall o he nex
memo y in he sequence [9].
The hippocampus, in conjunc ion wi h EC, p esen s an a chi ec u e and
unc ionali y wi h g ea po en ial o sol e he p oblem o he need o memo ies
wi h la ge capaci ies and low powe consump ion, as well as ha ing g ea appli-
cabili y in nume ous si ua ions. Howe e , e e ences and p oposals o his ype
o neu omo phic sys ems a e sca ce and limi ed in he li e a u e.
Wo ks such as [10, 11] p esen spike-based bio-inspi ed sequen ial memo y
models o he hippocampus, howe e , hey ha e a low s o age capaci y bo h in
e ms o numbe o memo ies and size o memo ies. Mo eo e , hey do no wo k
well wi h non-o hogonal memo ies (memo ies whose ac i i y con ains ac i a-
ions o neu ons in common).
The e a e o he al e na i es such as [12–15] in which neu omo phic memo y
models a e p esen ed, bu hey a e no sequen ial. In addi ion, [12,14,15] ocus
Neu omo phic bio-inspi ed sequen ial memo y model 3
mainly on demons a ing how hey wo k, so hey ha e a low s o age capaci y
and a e no able o wo k wi h non-o hogonal memo ies.
O he p oposed models o neu omo phic memo ies a e p esen ed in [16,17],
which a e no pu ely spike-based o [18,19] whose models a e no ully dynamic.
In iew o his sca ci y and limi a ions o al e na i es in he li e a u e, his
pape p oposes a spiking sequen ial memo y model bioinspi ed on he hippocam-
pus and EC o ma ion. The main con ibu ions o his wo k include he ollowing:
–A dynamic bio-inspi ed spiking sequen ial memo y model based on he hip-
pocampus and EC o ma ion, capable o lea ning and ecalling no only a
memo y, bu also sequences o memo ies.
–Implemen a ion o he p oposed SNN model on he SpiNNake ha dwa e
pla o m.
–The sou ce code is publicly a ailable, oge he wi h he documen a ion in-
cluding all he necessa y de ails ega ding he SNN a chi ec u es.
The es o he pape is s uc u ed as ollows: Sec ion 2 p esen s he ma e i-
als used in his wo k. In Sec ion 3, he p oposed model is de ailed, including i s
a chi ec u e (Sec ion 3.1) and i s ope a ion p inciple (Sec ion 3.2). The expe i-
men s pe o med o e alua e he unc ionali y and pe o mance o he p oposed
model a e explained in Sec ion 4, along wi h he esul s ob ained. Then, in Sec-
ion 5, he esul s o he expe imen s a e discussed. Finally, he conclusions o
he pape a e p esen ed in Sec ion 6.
2 Ma e ials
The model p oposed in his wo k has been buil using SNNs. These a e ne wo ks
ha a e no only bio-inspi ed by na u e, like A i icial Neu al Ne wo ks (ANNs),
bu also y o imi a e i . To do so, i consis s o 3 main compu a ional elemen s:
neu ons, synapses and lea ning ules.
Neu ons ha e a memb ane po en ial ha ep esen s hei in e nal s a e. This
po en ial is inc eased o dec eased by he a i al o asynch onous spikes om
o he neu ons ac oss synapses. When he memb ane eaches a h eshold, he
neu on gene a es a spike which i ansmi s ac oss he ou pu synapses. O all
he neu on models, he one mos widely used in he li e a u e is he Leaky
In eg a e-and-Fi e(LIF) [20].
O all he exis ing lea ning mechanisms, he Spike- iming-dependen plas ic-
i y (STDP) lea ning mechanism s ands ou . This is a Hebbian lea ning mecha-
nism in which he weigh o synapses is modi ied in p opo ion o he deg ee o
empo al co ela ion be ween he ac i i y o p e- and pos -synap ic neu ons [21].
The model p oposed in his wo k was implemen ed on he special-pu pose
ha dwa e pla o m o SNN a chi ec u es SpiNNake [22].
4 Daniel Casanue a-Mo a o e al.
Fig. 1. A chi ec u e o he p oposed sequen ial memo y model bio-inspi ed on he
Hippocampus and EC o ma ion.
3 Hippocampus compu a ional memo y model
3.1 A chi ec u e
The a chi ec u e o he p oposed sequen ial memo y model, which is inspi ed by
he hippocampus and EC, is p esen ed in Figu e 1.
The a chi ec u e is based on he model p oposed in [13]. This model p esen s
he base s uc u e o he hippocampus, which has been ex ended wi h he in-
oduc ion o EC laye s 2 and 5, as well as he connec ions and con igu a ion
necessa y o p ope and s able ope a ion.
As in he biological model, he hippocampus consis s o 3 popula ions o
neu ons: DG, CA3 and CA1.
DG consis s o a laye o neu ons in cha ge o inc easing he dispe sion o he
inpu in o ma ion. Fo his pu pose, a pa ial maximum spa si y o he inpu
in o ma ion was chosen. This popula ion o neu ons ac s as a decode and ca ies
ou a one-ho encoding o a pa o he inpu memo y, he es o he memo y
emaining unchanged. The encoded pa is called he cue o he memo y, and he
emaining unmodi ied pa is called he con en o he memo y. This unc ionali y
is achie ed hanks o he dis ibu ion o exci a o y and inhibi o y inpu synapses,
as well as he use o ecu en inhibi o y synapses ha egula e he oscilla ions
o he neu ons ha make up DG.
CA3 consis s o wo popula ions o neu ons, CA3cue o which he memo y cue
a i es and CA3con o which he memo y con en a i es. The CA3cue neu ons
a e connec ed ia synapses wi h he STDP lea ning mechanism o he CA3con
neu ons, ollowing an all- o-all connec i i y. This s uc u e is esponsible o he
lea ning and ecall o memo ies in he ne wo k.
CA1 has a s uc u e simila o DG, i.e., a laye o neu ons whose dis ibu ion
o exci a o y and inhibi o y inpu synapses de e mines he ou pu ac i i y. This
Neu omo phic bio-inspi ed sequen ial memo y model 5
s uc u e is esponsible o ca ying ou he e e se ope a ion o DG o eco e
he o iginal o ma o he memo y, i.e. i ac s as an encode . To do his, i
e-encodes he memo y cue and keeps he con en o he memo y as i is.
Wi h hese componen s, he minimal unc ionali y o lea ning, ecalling and
o ge ing indi idual memo ies is achie ed o o hogonal and non-o hogonal
memo ies. I is he in oduc ion o EC ha allows hese ope a ions o be ca ied
ou wi h sequences o memo ies.
The model p esen s EC laye s 2 (ECl2) and 5 (ECl5). ECl2 is esponsible o
ecei ing inpu in o ma ion in o he ne wo k and passing i o he hippocampus.
ECl5 has a dual unc ion, o ake he memo y ou o he hippocampus and send
pa o i back o he inpu , speci ically o ECl2.
The whole memo y is no sen back o ECl2, bu only pa o i s con en .
On he one hand, ECl5 has exci a o y synapses ha connec he neu ons whose
ac i i y encodes he pa o he con en o be sen back o he inpu wi h he
neu ons in ECl2 whose ac i i y will encode he cue o he inpu memo y. On
he o he hand, ECl5 also has inhibi o y synapses ha connec he neu ons ha
encode he cue o he ou pu memo y wi h he ECl2 neu ons ha will encode
he cue o he inpu memo y.
In sho , he synapses connec ing ECl5 and Ecl2 o m a second ecu en
ne wo k esponsible o gene a ing a sequen ial ecall o lea ned and s o ed mem-
o ies.
3.2 Ope a ing p inciple
The spiking ac i i y o he model p esen s a spa ial coding in ime. This encoding
is based on he ac ha all neu onal ac i i y wi hin a subpopula ion a he same
ins an o ime e e s o he same memo y, while neu onal ac i i y in he same
subpopula ion bu a di e en ins an s o ime e e s o di e en memo ies.
Thus, he p oposed model is able o pe o m he ope a ions o lea ning,
indi idual and sequen ial ecall o memo ies and o ge ing.
Lea ning ope a ion s a s when a comple e memo y a i es a he inpu o
he ne wo k (ECl2) and is main ained o 3 ime uni s. Only he ac i i y o he
i s and hi d ime uni will gene a e in e nal ac i i y in he ne wo k. Tha is
because, du ing he second ime uni , he neu ons will be in e ac o y pe iod
due o he ac i a ion o he i s ime uni .
Pa o his memo y ( he cue), will be encoded as i passes h ough DG and
bo h pa s will a i e a he same ime a CA3. When hey a i e a he same
ime, he STDP will be ac i a ed, p oducing an inc ease in hose synapses whose
p e- and pos -synap ic neu ons a e ac i a ed. In o he wo ds, lea ning o he
inpu memo y akes place. This memo y will pass o CA1 whe e i will eco e
he inpu o ma and lea e he ne wo k h ough ECl5. Thanks o inhibi o y
synapses om ECl5 o ECl2, du ing he lea ning ope a ion, sequen ial ecall will
no be ac i a ed. The ope a ion equi es a o al o 7 uni s o ime o comple e.
Recall ope a ion s a s when a agmen o he memo y o be ecalled (cue)
a i es a he ne wo k inpu du ing a single ime uni . I will a i e a CA3cue
a e being encoded in DG. The ac i a ion o neu ons in CA3cue will p oduce
6 Daniel Casanue a-Mo a o e al.
he ac i a ion o a se o neu ons in CA3con whose synapses ha e p e iously
been inc eased in weigh wi h a lea ning ope a ion. In o he wo ds, a ecall o
he memo y cha ac e ized by he cue cue will be p oduced. This memo y will
each CA1 whe e i will e ie e he inpu o ma and lea e he ne wo k h ough
ECl5. The ope a ion equi es a o al o 6 uni s o ime o comple e.
This would be he case o he ecall o a single memo y. I he ac i i y o he
memo y ha is passed back o ECl2 con ains some in o ma ion, i would en e
he hippocampal model again as a cue o s a he ecall o he nex memo y
in he sequence. Tha is why he pa o he con en ha is passed back o he
ne wo k o s a he nex ecall ope a ion is called nex cue. Fu he mo e, by
egula ing he delay o he synapses connec ing ECl5 and ECl2, he equency
o ecall o his sequence o memo ies can be modi ied.
Finally, he o ge ing ope a ion is pe o med implici ly du ing he lea ning
ope a ion when he inpu memo y ( o be lea ned) has a cue in common wi h
ano he p e iously lea ned memo y. A he same ime ha a lea ning ope a ion
is p oduced on he new memo y, a ecall ope a ion o he p e ious memo y is
p oduced as a consequence, since hey use he same cue. Howe e , due o he
empo al sequence in which hese ope a ions occu , he STDP causes an inc ease
in he synapses in ol ed in he lea ning o he new memo y and a dec ease in
he synapses in ol ed in he ecall o he p e ious memo y. In o he wo ds, a
lea ning o he new memo y and o ge ing o he old one.
4 Expe imen a ion and esul s
In o de o e i y he co ec ope a ion o he ne wo k, a se o wo expe imen s
ha e been ca ied ou . These expe imen s ha e been ca ied ou in eal- ime and
ha e been applied on he model implemen ed in SpiNNake whose ime uni is
1 milisecond (ms).
Each expe imen consis ed o a empo al simula ion o he model in which
he beha io o he ne wo k is analyzed gi en an inpu ac i i y. The esul s
o each expe imen a e p esen ed g aphically by means o as e plo s. The X-
axis ep esen s he empo al e olu ion o he simula ion in milliseconds and he
Y-axis ep esen s each neu on o he ne wo k iden i ied by he subpopula ion
o which i belongs and he in e nal ID wi hin he subpopula ion. Each poin
ep esen s a spike i ed by he neu on ma ked by he Y-axis a he ime ins an
ma ked by he X-axis.
Fo bo h expe imen s, a ne wo k wi h a s o age capaci y o up o a maximum
o 6 simul aneous memo ies, wi h a size o 16 neu ons o encodable in o ma ion
ac i i y each, was used. O hese 16 memo y neu ons, he i s 3 encode he
cue and he nex 13 he con en . O he 13 neu ons whose ac i i y encodes he
con en , he las 3 encode he nex cue. A delay o 50 ms has been used o ecu -
en synapses connec ing ECl5 o ECl2. Ne e heless, he p oposed model is ully
pa ame e ised and can be used o smalle o la ge capaci y implemen a ions.
The i s expe imen pe o ms a lea ning and ecall ope a ion on one memo y,
ollowed by ano he pai o lea ning and ecall ope a ions on a second memo y.
Neu omo phic bio-inspi ed sequen ial memo y model 7
Fig. 2. Ras e plo o spiking ac i i y o he ne wo k du ing he ope a ion es consis -
ing o lea ning, ecall and o ge ing ope a ion.
Bo h memo ies ha e in common he cue and a pa o he con en . This expe i-
men a emp s o demons a e he ope a ion o he basic ope a ions o lea ning,
ecalling and o ge ing, as well as hei abili y o ope a e on non-o hogonal
memo ies.
The expe imen had a du a ion o 27 ms and he ne wo k ac i i y as a esul
o he expe imen can be seen in Fig. 2. I begins wi h he lea ning ope a ion
(ms 0 o ms 8) o he memo y o med by he ac i a ion o neu ons IN0, IN1, IN3,
IN4 and IN5. Once lea ned, i mo es on o a ecall ope a ion (ms 7 o ms 14) by
in oducing he cue o his i s memo y, IN0 and IN1. A e a ew oscilla ions,
he comple e memo y can be seen a he ou pu o he ne wo k a ms 13 (cue)
and ms 14 (con en ). This shows ha he memo y has been co ec ly lea ned
and ecalled.
Nex , lea ning (ms 13 o ms 21) o he second memo y o med by he ac i a-
ion o neu ons IN0, IN1, IN5, IN6 and IN7 begins. Du ing he lea ning p ocess
o he new memo y in CA3 (ms 16 o ms 18), he ecall o he p e ious memo y
8 Daniel Casanue a-Mo a o e al.
Fig. 3. Ras e plo o spiking ac i i y o he ne wo k du ing he ope a ion es consis -
ing in he lea ning o a sequence and comple e ecall o his sequence.
can be seen. A e ha , he memo y cue ( he ac i a ion o IN0 and IN1 neu ons)
is in oduced in o he ne wo k o s a he ecall ope a ion (ms 20 o ms 27).
A e he ecall, he ou pu ac i i y o he ne wo k indica es ha he second
memo y has been co ec ly ecalled and he i s has been o go en.
The second expe imen pe o ms lea ning ope a ions o a sequence o memo-
ies, ending wi h a ecall ope a ion o he whole sequence. This expe imen makes
i possible o demons a e he unc ioning o memo y wi h sequen ial memo ies.
The expe imen has a du a ion o 138 ms and he ac i i y o he ne wo k
as a esul o he expe imen can be seen in Fig. 3. I s a s wi h 3 lea ning
ope a ions (ms 0 o ms 8, ms 7 o ms 15 and ms 14 o ms 22). The i s memo y
is o med by he ac i a ion o neu ons IN0, IN3, IN4, IN5 and IN14, he second
memo y by he ac i a ion o neu ons IN1, IN4, IN5, IN6, IN13 and IN14, and
he hi d memo y by he ac i a ion o neu ons IN0, IN1, IN5, IN6, IN7.
Nex , a ecall ope a ion is pe o med a ms 21, using he ac i a ion o he
IN0 neu on as he cue. This ope a ion succeeds in co ec ly ecalling he i s
Neu omo phic bio-inspi ed sequen ial memo y model 9
memo y a ms 28. A he same ime, he ac i i y o he las 3 neu ons is shi ed
om ECl5 o ECl2. Tha is, IN14 in ECl5 will each ECl2 as IN1 50 ms la e .
This will esul in he beginning o he ecall ope a ion o he second memo y
(ms 77). A ms 83 he con en s o he second memo y will be ou pu om he
ne wo k, whose neu ons IN13 and IN14 will pass as cues (IN0 and IN1) o ECl2,
and he hi d ecall ope a ion will begin (ms 132). Finally, a ms 138, he con en
o he hi d memo y is ou pu ed om he ne wo k. This memo y has no ac i i y
in he las h ee neu ons, so he ecall sequence inishes.
5 Discussion
The esul s o he expe imen s de ined in Sec ion 4 demons a e he co ec unc-
ioning and ope a ional capabili ies o he p oposed model: lea ning, indi idual
and sequen ial ecall o memo ies and o ge ing.
The p oposed model is no only inspi ed by, bu also ies o mimic he
biological model o hippocampal o ma ion and EC. This is why he p oposed
model consis s o a hippocampal s uc u e o med by 3 main laye s: DG, CA3
and CA1, wi h unc ionali ies simila o hose p esen in he biological model.
The hippocampus and EC model oge he p o ide he sys em wi h he abili y o
lea n, ecall indi idual and sequen ial memo ies and o ge , as in he biological
model. Finally, o all he EC laye s, laye s 2 and 5 ha e been used o assemble
he minimal ecu en sys em capable o ca ying ou his sequen ial ecall.
Rega ding he models p oposed in he li e a u e, he model desc ibed in
his pape is compa ed wi h hose men ioned in Sec ion 1. Unlike he models
p esen ed in [10,11], he model p oposed in his wo k is able o wo k wi h bo h
o hogonal and non-o hogonal memo ies and has a la ge s o age capaci y. In
[12–15], models p esen ed a e only capable o pe o ming single ecall ope a ions,
bu lack he possibili y o linking se e al memo ies sequen ially and accessing
hem in his way, as is possible in he model p oposed in his pape .
Finally, in [16,17] he models a e no pu ely spike-based and in [18,19] a e no
ully dynamic, whe eas he p oposed model is pu ely spike-based and dynamic.
This ansla es in o lowe esou ce consump ion and be e scalabili y wi h an
app oach ha is close o he biological model.
6 Conclusions
In his wo k, a ully- unc ional bio-inspi ed sequen ial model o he hippocampus
and EC was p oposed and implemen ed wi h SNNs on he SpiNNake ha dwa e
pla o m. The model is capable o pe o ming lea ning, indi idual and sequen ial
ecall o memo ies and o ge ing ope a ions. To demons a e he co ec unc-
ioning o he model and i s di e en ope a ions, a se o expe imen s based on
ope a ional es s we e pe o med.
This expe imen a ion has been de eloped o a conc e e implemen a ion o
he model. Howe e , hanks o he pa ame e iza ion wi h which he model is
de ined, i also wo ks o implemen a ions o g ea e o lesse s o age capaci y,
