Noise h ough o wos and se ens: c ea ing audio isual
a wo ks om a i icial neu al ne wo ks’ p ocessing
da a
Tanguy Pocque
Depa men o Music
Uni e si y o Manches e
B idge o d S , Manches e M13 9PL
[email p o ec ed]
Abs ac
This piece is he esul o he soni ica ion and isualiza ion o an a i icial neu al
ne wo k’s p ocessing da a. A i s co e, i is an a emp a making a om a i icial
in elligence wi h a non-gene a i e app oach, lea ing all c ea i e agency o he a is ;
a explo ing s uc u es inhe en o he ope a ion o neu al ne wo ks; a commen ing
on he na u e o hese ubiqui ous, ye no o iously unin elligible algo i hms. I does
so by ocusing on some o he ne wo k’s in e nal da a s eams du ing aining, look-
ing a hei g adual mo emen om noise o o de , and mapping his in o ma ion
on o sound and isual ma e ials and p ocesses de i ed om ‘gli ch’ and d one
adi ions, ha o e in e es ing aes he ic pa allels o his e olu ion o he da a.
1 Technical app oach and mo i a ions
A i icial neu al ne wo ks (NNs) a e commonly known as ‘black box algo i hms’. Ex ensi e wo k
has been done on he need o emedy he ac ha , gene ally, “no in o ma ion is p o ided abou
wha exac ly makes [neu al ne wo ks] a i e a hei p edic ions” (Samek e al., 2017), and coun less
app oaches ha e been aken o achie e his (Adadi and Be ada, 2018; Ali e al., 2023). F om a
echnical s andpoin , his wo k si ua es i sel in his sea ch o means o make a i icial in elligence
(AI) mo e ‘in e p e able’ (p o iding a “le el o unde s anding o how he unde lying echnology
wo ks” (ISO/IEC, 2020)).
Ra he han use he p oduc o a ained neu al ne wo k (a gene a i e app oach), i looks a he p ocess
ha enables his ou pu , a da a and me ics om aining, when billions o pa ame e s a e i e a i ely
weaked, un il he desi ed ou pu is eached o any gi en inpu . These da a s eams (common
pe o mance me ics such as loss and accu acy; and syn he ic me ics aken om he gene a ed da a
such as en opy a e ages o he samples, simila i y be ween samples, RMS alues, edge in ensi y,
e c.) a e hen mapped on o sound pa ame e s ( a ious g anula syn hesize s’ pa ame e s, sample
leng hs, o e d i e, o e all noise, e c.) and isuals. Beyond o e ing some insigh s in o how NNs
ope a e, he hope is o p esen an al e na i e o gene a i e app oaches in he making o a wo ks based
on — o commen ing on — AI.
This piece uses a gene a i e ad e sa ial ne wo k (GAN), ained on he MNIST da ase o images o
handw i en digi s. The use o as commonplace and simple a da ase as possible el essen ial o keep
he ocus on he p ocess, on he e olu ion o he da a p esen ed o e he cou se o he aining p ocess.
This piece illus a es how GANs ain om noise o o de , showcasing a ious possible in e media y
s uc u es.
The non- eleological app oach is made clea e in he di e ences be ween he p io i ies o mos NN
designe s, and ha o his wo k. The la e is o c ea e aes he ically and na a i ely in e es ing esul s,
including h ough he p esen a ion o sub-op imal uns, he aming o inaccu acies, o he in en ional
push owa ds ine iciency. The idea is no o c ea e a pa icula ly e icien GAN, bu o demons a e
P oceedings o he 6 h Con e ence on AI Music C ea i i y (AIMC 2025),
B ussels, Belgium, Sep embe 10 h-12 h
how i ains, and some imes ails. Some o he uns displayed in he piece a e examples o mode
collapse, unde i ing, o ou igh andom ou pu om he ne wo k, all o which a e ypical ailu es
om a ne wo k design pe spec i e. The s ance aken being ha he e is a lo o unde s anding o
hese algo i hms o gain om seeing hem ail.
This app oach ies in wi h he main aes he ic cu en s ha he musical and isual sides o he wo k
a e indeb ed o: ‘pos -digi al’ and ‘gli ch’ adi ions.
2 Aes he ic and a is ic con ex
While echnically algo i hmic soni ica ion, his wo k links o sys ema ic music as a whole, algo i hmic
music mo e speci ically, and e en mo e speci ically AI app oaches o algo i hmic music (Todd, 1989).
I also links o o he examples o NNs being epu posed o non-gene a i e music-making (Tudo ,
1995).
In essence, e lec ing some o he echnical p io i ies de ailed abo e, his wo k is pe haps an illus a ion
o he iew ha “The poin o a is de i us” (Weine , 2012); i ocuses on p ocesses, and on wha a e
no mally byp oduc s, ex aneous ma e ial.
Sonically, i owes a lo o ‘gli ch’ music, a subgen e ha ha nesses “ he ‘ ailu e’ o digi al echnology
[. . . ]: gli ches, bugs, applica ion e o s, sys em c ashes, clipping, aliasing, dis o ion, quan iza ion
noise, and e en he noise loo o compu e sound ca ds a e he aw ma e ials compose s seek o
inco po a e in o hei music” (Cascone, 2000). Mo e b oadly, i links o he pos -digi al aes he ic,
which “ akes digi al o be he pinnacle o audio ideli y ( echnology d i en by he ul ima e desi e o
pe ec cla i y and he elimina ion o noise) and si ua es i sel a e : a e he e olu ion, amids he
edisco e y o noise h ough digi al mal unc ion” (Hawo h, 2016). This link be ween he in o ma ion
being soni ied (byp oduc s, de i us) and he sound hey e ec (sounds o e o s, digi al mal unc ion,
clicks, and noise) is cen al o he p ojec .
In many da a-d i en pieces, he e seems o be a need o au al ep esen a ions o pa sing; scanning;
il e ing; g ouping; o de ing; e c., all concep s we associa e wi h da a-handling. This can ake he
o m o sounds associa ed wi h ba code scanne s (Ikeda, 2008, 2005); ha d d i es’ ead-and-w i e
heads (Panacea’s Syneco e); elec omagne ic gli ch/clicks/in e e ence (Ikeda, 1996; SND, 2010);
a human oice eading ou li e al da a/nume ical in o ma ion (al a no o, 2008). This ex ends o
isual/seman ic ep esen a ions o da a: ba codes (Ikeda, 2005); bina y numbe ing (Ikeda, 2008);
e e ences o coding (SND, 2000); ma hema ical concep s (Ikeda, 1996); o a sel - e e en ial ocus
( he music i sel as he da a) (Goem, 2013; SND, 2000; Cascone e al., 2004; Cyclo, 2001). This
appa en need o links o ma e ial p ocesses in au al ep esen a ions o da a ha is — by na u e —
imma e ial, is also some hing his piece a emp s o add ess.
As Eno obse ed, he use o hese ypes o sounds, and o hei seman ic implica ions, is bo n o a
need o ma e iali y in elec onic music: “a e he loss o ma e iali y wi h he appea ance o digi al
music, one migh see he ad en o music e oking imagined bu speci ic and angible media as caused
by a need o ema e ialise sound” (Weium and Boon, 2013). This piece migh lead o one o hese
imagined media, linked o, bu dis inguishable om, he ac ual algo i hmic p ocesses i is d i en by.
This piece also uses NNs as ano he c ea i e ool o sound s uc u ing (Hawo h, 2016). I explo es
he idea ha he as amoun s o da a gene a ed by mode n echnology o m “a can as o he
con empo a y a is ” (Lynch and Pa adiso, 2016), leading o a “da abase aes he ic” (Vesna, 2007).
Finally, i displays ma e ials linked o ins umen al and elec onic d one adi ions (Da achi, 2023;
Malone, 2019; Akb o, 2017; Radigue, 2018; Con ad, 2017). A lo o he sou ce ma e ials a e o gan
eco dings, o a ious d ones and imp o isa ions. The con as be ween he highly syn he ic gli ch
sounds, and highly o ganic and inhe en ly human sounds o a chu ch o gan o ms an in e es ing
pa allel wi h he GAN’s g adual ep oduc ion o handw i en digi s, wi h he mo e om highly
abs ac ed, andom nume ical da a, o an in elligible, human-looking ou pu .
Re e ences
Adadi, A. and Be ada, M. (2018). Peeking inside he black box; a su ey on explainable a i icial
in elligence. 6:52138–52160.
2
Akb o, E. (2017). Fo o gan and b ass. Sub ex eco dings, London.
Ali, S., Abuhmed, T., El-Sappagh, S., Muhammad, K., Alonso-Mo al, J., Con alonie i, R., Guido i,
R., Del Se , J., Diaz-Rod iguez, N., and He e a, F. (2023). Explainable a i icial in elligence
(XAI): Wha we know and wha is le o a ain us wo hy a i icial in elligence. 99(101805).
al a no o (2008). uni x . Ras e -No on, Chemni z.
Cascone, K. (2000). The aes he ics o ailu e: "pos -digi al" endencies in con empo a y compu e
music. 24(4):12–18.
Cascone, K., Kahn, J., and s einb uechel (2004). ATAK004. ATAK, Tokyo.
Con ad, T. (2017). Ten yea s ali e on he in ini e plain. Supe io Viaduc , San F ancisco.
Cyclo (2001). .NOTON, Chemni z.
Da achi, S. (2023). Long g adus: a angemen s. La e music, Los Angeles.
Goem (2013). S ud s im. Ko m digi aal, Nijmegen.
Hawo h, C. (2016). ‘all he musics which compu e s make possible’: ques ions o gen e a he p ix
a s elec onica. 21(1):15–29.
Ikeda, R. (1996). +/-. Touch, London.
Ikeda, R. (2005). da aplex. Ras e -No on, Chemni z.
Ikeda, R. (2008). es pa e n. Ras e -No on, Chemni z.
ISO/IEC (2020). So wa e and sys ems enginee ing, so wa e es ing, pa 11: Guidelines on he
es ing o AI-based sys ems.
Lynch, E. and Pa adiso, J. (2016). Senso Chimes: Musical mapping o senso ne wo ks. B isbane,
Aus alia.
Malone, K. (2019). The sac i icial code. Ideologic o gan, Pa is.
Radigue, E. (2018). Occam ocean I. Shiiin, Pa is.
Samek, W., Wiegand, T., and Mülle , K.-R. (2017). Explainable a i icial in elligence; unde s anding,
isualizing and in e p e ing deep lea ning models.
SND (2000). s dio. Mille pla eaux, F ank u .
SND (2010). makesnd casse e. Mille pla eaux, F ank u .
Todd, P. (1989). A connec ionis app oach o algo i hmic composi ion. 13(4):27–43.
Tudo , D. (1995). The neu al ne wo k syn hesize .
Vesna, V. (2007). Da abase aes he ics - a in he age o in o ma ion o e low. Uni e si y o
Minneso a p ess, Minneapolis.
Weine , L. (2012). Law ence weine in e iewed by hans ul ich ob is .
Weium, F. and Boon, T. (2013). Ma e ial cul u e and elec onic sound. A e ac s: s udies in he
his o y o science and echnology. Smi hsonian ins i u ion schola ly p ess, Washing on, D.C.
3
