Fading Faces: When Agen s Fo ge Who You A e
Da id F ei e-Ob eg´
on 1
1SIANI, Uni e sidad de Las Palmas de G an Cana ia, Spain
da[email p o ec ed]
Abs ac
As a i icial agen s o m hei cogni i e depic ion o o he s,
hei isual ecogni ion o wha hey once knew becomes
agile. In his wo k, we discuss pe cep ual o ge ing ia a
minimalis simula ion o embodied agen s. Agen s ha e dis-
inc isual iden i ies ( aces) and pe sonal con olu ional neu-
al ne wo ks (CNN) lea ned online o iden i y neighbo s. Ini-
ially, aces appea in high de ail, bu g adually decay o min-
imal igu es, modeling cogni i e aging o pe cep ual decline.
We see a p ecipi ous d op in ecogni ion accu acy while sym-
bolic iden i y emains in ac , illus a ing a dissocia ion be-
ween knowing who a pe son is and ecognizing pe sons. Ou
esul s place o ge ing no as e asu e bu as pe cep ion ade, a
symbolic-pe cep ual dissocia ion esonan wi h ac ual cogni-
i e mo ali y, p omp ing conce ns abou embodimen , mem-
o y, and decay in a i icial li e.
In oduc ion
Recogni ion is a basic ype o social memo y (Wang and
Zhan, 2022). Fo humans and machines alike, knowing
who a pe son o objec is depends upon a enuous chain o
pe cep ual, symbolic, and embodied p ocedu es. Howe e ,
whe eas machine ecogni ion ends o p esuppose ideal,
noise- ee si ua ions, eal pe cep ion de e io a es, ans-
mu es, o disappea s al oge he . We model pe cep ual o -
ge ing wi h a minimal simula ion whe e agen s on a 2D g id
each ha e a dis inc ace. Using a ine- uned ResNe 18, hey
lea n o ecognize neighbo s du ing encoun e s. O e ime,
aces deg ade om de ailed images o ske ches, e oding pe -
cep ion: iden i ies emain unchanged, ye ecogni ion ails.
Ou se up d aws i s impulse om cogni i e heo ies o mo -
ali y. Fo ge ing is no dicho omic in biological li e: one
can ecall a name bu no a ace, o ecall an emo ion bu no
he indi idual. Simila ly, in ou agen s, symbolic iden i y
(an un o ge able, unique ID) i sel is e ained, bu pe cep-
ual access o such iden i y decays. The e ec is a de e io a -
ing o m o mo ali y, mo e wo ied abou e osion o amil-
ia i y han abou anishing. This p ocess alls wi hin con-
igu a ions wi hin au opoiesis (F oese e al., 2023), home-
odynamic sys ems (Ya es, 1994), and ep esen a ional d i
(D iscoll, 2022), whe e in e nal models dissocia e om ex-
300 S okes 150 S okes 75 S okes 30 S okes
Figu e 1: Snapsho o he 5×5 agen en i onmen , wi h i e
agen s andomly dis ibu ed in he g id. A he bo om, wo
sample iden i ies a e shown ac oss ou abs ac ion le els
(300, 150, 75, and 30 s okes), illus a ing he isual deg a-
da ion p ocess. These inc easingly abs ac depic ions simu-
la e he memo y decay agen s ace when ecognizing o he s
unde ha she pe cep ual condi ions.
e nal eali y. Dissocia ion, he e, does no o igina e om
an agen ’s body, bu om pe cep ion by a body o o he s.
We he eby show how minimal agen s, equipped wi h neu al
pe cep ion and sho - e m memo y, can be exposed o sym-
bolic decay.
Expe imen al Se up
We implemen an agen -based simula ion using he Mesa
amewo k. Each agen is assigned an ini ial acial image,
andomly selec ed om he KDEF da ase , es ic ed o neu-
al exp essions and one image pe iden i y (Lundq is e al.,
1998). Faces a e ende ed wi h a ying deg ees o abs ac-
0
20
40
60
150 s okes
80
75 s okes
100
30 s okes
120
S ep
0.30
0.35
0.40
0.45
0.50
0.55
0.60
A g. Recogni ion Con idence
Figu e 2: A e age ecogni ion con idence o agen s o e ime. Red lines indica e deg ada ion s eps.
ion using CLIPasso ske ches (Vinke e al., 2022), o ou
le els: 300, 150, 75, and 30 s okes. The 300-s oke en-
de ing is he mos de ailed image; low-s oke ende ings
p og essi ely deg ade isual cla i y (F ei e-Ob eg´
on e al.,
2025). Agen s andomly mo e inside a 2D o oidal g id.
Upon en e ing neighbo ing cells, wo agen s a emp o ec-
ognize each o he . They ecognize each o he based on a
CNN (ResNe 18), and each agen has i s own ne wo k. Du -
ing he i s 70 s eps, agen s inc emen ally lea n: whene e
hey see an un amilia ace, hey assign i a new class la-
bel and upda e hei ne wo k (only he inal, i.e., classi ica-
ion, laye is ained). Agen s s op lea ning om s ep 71 and
begin e alua ing ins ead. A ce ain in e als, acial isual
quali y de e io a es ( om 300 o 150, o 75, o 30 s okes).
The iden i y o he agen ne e a ies, bu i becomes ha de
o classi y i s pe cep ual da a. To model he ecogni ion de-
cay, we keep ack o whe he o no agen A co ec ly clas-
si ies agen B when hey encoun e each o he . I he p edic-
ion co esponds o he ac ual label lea ned a aining, we
say ecogni ion has succeeded; i no , we ma k i as a ail-
u e. Each agen moni o s i s accu acy o e ime and e eals
a g adual decline in ecogni ion pe o mance as isual da a
decays. T us in agen s, measu ed as co ec p edic ions di-
ided by he numbe o a emp s, also a ies wi h hei isual
size, signi ying he le el o con idence. Such deg ada ion
mi o s how humans nego ia e iden i y wi h pa ial memo y,
showing o ge ing as pa o social in e ac ion a he han a
law.
Resul s and Analysis
The de elopmen o o e all a e age us o all agen s is
p esen ed in Figu e 2, beginning om s ep 40, once assess-
men commences. Fo he ini ial 40 s eps, agen s ga he no
isual knowledge o he han h ough local in e ac ion; no
measu e o ecogni ion pe o mance is ye conduc ed.
Once e alua ion s a s, we obse e a apid ise in a e -
age us , eaching a peak nea s ep 70. This indica es ha
agen s success ully consolida e hei ecogni ion capabili ies
du ing he lea ning phase, pa icula ly when he inpu im-
ages main ain high ideli y (300-s oke ske ches). Howe e ,
om s ep 70 onwa ds, when he ske ches g adually de e io-
a e, us le els s a o come down. The decline accele a es
when agen s mo e om 150-s oke o 75-s oke and inally
o 30-s oke ep esen a ions. A s ep 130, a e age us le -
els d op o less han 0.35, a se e e deg ada ion o ecogni-
ion accu acy. This beha io a o s ou hypo hesis: agen s
o ge no by dele ing in e nal ep esen a ions, bu by pe -
cei ing pe cep ual misma ch be ween s o ed ep esen a ions
and newly co up ed inpu s. Al hough s ill main aining hei
lea ned CNN pa ame e s, agen s canno ma ch mo e abs ac
ske ches, leading o a symbolic o m o o ge ing.
Conclusions
This wo k p o ides a symbolic accoun o memo y and o -
ge ing in embodied agen s h ough ace-based social in e -
ac ion. We show how agen s, wi h localized CNNs and ini e
li espans, acqui e ecogni ion abili ies ia si ua ed in e ac-
ion, bu g adually de e io a e, as pe cep ual ideli y wo s-
ens. Impo an ly, o ge ing in his scena io does no a ise
due o e asu e om memo y bu due o a misma ch be ween
in e nal ep esen a ions and ex e nal inpu s, ins an ia ing a
s uc u al unde s anding o o ge ing. These aise ques ions
abou how pe cep ual decline and symbolic loss shape no
only a i icial memo y sys ems, bu also he eme gence o
sociali y and cogni i e de elopmen , sugges ing ha mo al-
i y in agen s mi o s he human condi ion o emembe ing
and o ge ing.
Acknowledgemen s. This wo k is pa ially unded
unded by p ojec PID2021-122402OB-C22/MICIU/AEI
/10.13039/501100011033 FEDER, UE and by he ACIISI-
Gobie no de Cana ias and Eu opean FEDER unds unde
p ojec ULPGC Facili ies Ne and G an EIS 2021 04.
Re e ences
D iscoll, L. N. (2022). Rep esen a ional d i : Eme ging heo-
ies o con inual lea ning and expe imen al u u e di ec ions.
Cu en Opinion in Neu obiology, 76:102609.
F ei e-Ob eg´
on, D., Ne es, J., ˇ
Ziga Eme ˇ
siˇ
c, Meden, B., Cas-
ill´
on-San ana, M., and P oenc¸a, H. (2025). Syn hesizing
mul ile el abs ac ion ea ske ches o enhanced biome ic
ecogni ion. Image and Vision Compu ing, 154:105424.
F oese, T., Webe , N., Shpu o , I., and Ikegami, T. (2023). F om
au opoiesis o sel -op imiza ion: Towa d an enac i e model
o biological egula ion. bioRxi .
Lundq is , D., Flyk , A., and ¨
Ohman, A. (1998). The ka olinska
di ec ed emo ional aces (kde ). CD-ROM om Depa men
o Clinical Neu oscience, Psychology sec ion, Ka olinska In-
s i u e . ISBN 91-630-7164-9.
Vinke , Y., Pajouheshga , E., Bo, J. Y., Bachmann, R. C., Be mano,
A. H., Cohen-O , D., Zami , A., and Shami , A. (2022). CLI-
Passo: Seman ically-Awa e Objec Ske ching. ACM T ans.
G aph., 41(4).
Wang, X. and Zhan, Y. (2022). Regula ion o social ecogni ion
memo y in he hippocampal ci cui s. F on ie s in Neu al Ci -
cui s, 16:839931.
Ya es, F. (1994). O de and complexi y in dynamical sys ems:
Homeodynamics as a gene alized mechanics o biology.
Ma hema ical and Compu e Modelling, 19(6):49–74.
