0:13

hi hi everybody thanks for coming and joining us in the very last session of today

0:19

thanks ryan for hosting us all as part of the conference

0:24

so my name is nora margu i’m a curator and based out of ireland i’m interested in

0:30

i guess the the creative expression around

0:35

artificial intelligence and how it can call attention to those complexities of like the social political conditions in

0:41

which algorithms and computation resides in within our everyday um

0:47

for this panel we are looking at machine intelligence in the context of art and we’re looking at the new types of

0:52

knowledge values and meanings that these technical modalities make possible and

0:58

how we might kind of read and understand artistic work in kind of contemporary technological conditions so i’m jo and

1:06

i’m i’m joined today by three different speakers whose practices um are across a broad

1:13

range of subjects and um uh the first speaker that we have today

1:19

is by ben um do you want me to give you a little introduction yeah

1:25

sure so ben bogart is a vancouver-based interdisciplinary artist working with generative computational processes

1:32

involving physical modeling chaos feedback systems and computer vision and machine learning and he’s been inspired

1:38

by knowledge in the natural science sciences including quantum physics and cognitive

1:43

neuroscience um i guess that’s an interesting place to

1:49

maybe leave it there for your bio and today he’s going to talk about machine subjectivity

1:55

thank you norah i forgot i had to pick up the second microphone hello everybody

2:01

uh so i am going to be talking about the subjective machines and part of this is supposed to be a little weird

2:08

and challenging so feel free to push back on everything so i’m interested in subjective machines

2:15

because of my work with machine learning and i’m going to call them learning machines because i’m interested in

2:21

thinking about them as objects and methods not necessarily disciplines

2:26

machine learning is a discipline and so learning machines are something that i’ve been working on

2:32

since about 2007 so they’ve been quite integrated in my practice

2:38

and i’ve been thinking about them in a particular way through exploring them in in various artistic

2:44

manifestations so i think it might be worthwhile to talk a little bit about how i think about

2:50

these machines that learn and part of that is because there’s just so much hype and mystification about

2:58

what they can and can do and so i would define learning machines as any system that can improve their

3:04

performance on some task autonomously often the conduction the performance of

3:11

the task involves some kind of statistical model sometimes the machine learning’s purpose is only to make a

3:17

statistical model but at least in the kinds of things like deep networks we’re talking about most often these days and

3:24

the leaps and bounds these technical systems are reaching those are almost always

3:30

exclusively being conducted by a statistically oriented

3:35

machine learning system so i want to really enunciate and emphasize that

3:41

the kinds of models constructed by learning machines are the same kinds of models that human

3:48

statisticians would make there’s nothing miraculous about learning machines they are not magic they are not necessarily

3:54

the future but they could be so why are they special if they’re just more statistical models of which we have

4:01

no shortage uh i think of kind of two polls one is a pro and the other one’s a

4:08

con the pro and the real reason why they’re so popular and we’re talking about them so much is the degree to which they’re

4:14

scalable we can make really really big statistical models trained on lots and

4:19

lots and lots of data now and we can do it in a way that’s much more cost effective than it would have been had we

4:25

done it with human statisticians even those using computers so the big part of why things are so

4:31

popular is that it’s cost effective it is has a value in a capitalist system to

4:36

autonomize even statistics but the con is to what degree they are

4:43

actually interrogable and by that i mean how much can we question their results

4:48

how can we have a conversation with them about why they have made a statistical conclusion

4:55

and they’re not very good at that especially the deep networks we’re talking about most often when we talk

5:01

about machine learning today so they’re really black boxes that are opaque and invisible and as they’re

5:08

deployed in all kinds of contexts through social media and marketing systems in a capitalist context we have

5:14

to be really careful about what they could be saying about us in terms of the models they’re building

5:19

and so for me one really big question is how do we validate these models if the only cost away

5:25

cost effective way to produce them is by building another giant machine learning

5:31

system that does the same thing if we can only compare them to each other because we can’t compare them to a model

5:37

constructed by a human how do we know they got it right especially when we can’t interrogate it

5:43

so that’s my entry point for subjective machines is kind of pushing against this idea of the

5:49

objectivity of the statistical model in the context of autonomous

5:54

machine learning systems so i’m not going to define well i am going

6:00

to define subjectivity but i’m going to find it in a very narrow way that comes out of my practice and thinking about learning machines so a subjective

6:07

machine subject to the idea of subjective machines is a framework for thinking

6:12

about objectivity and subjectivity both in autonomous machines and also in brains

6:20

that are in biological material and this kind of follows from my phd work which was funded by the social

6:27

science and research council of canada on making a machine dream and you can ask me more about that later i’m not

6:33

going to talk about it so i would i’m going to get the definition

6:38

out of the way right away and then talk more about details but i define subjectivity as an interaction between

6:44

sensation and imagination that forms a reinforcing pattern resulting in

6:49

perception simple enough there are two philosophical assumptions

6:55

behind this definition of subjectivity there’s probably more but these these are two that i’m picking out and these

7:01

ones are really important to me because they’re my own personal beliefs also the first is that the world independent of

7:07

cognition that is the world that isn’t that we cannot conceive of that is beyond senses and perception

7:14

is actually there but it is unevenly distributed and continuous

7:21

it’s not made up of discrete chunks of things although some physicists would disagree

7:27

and second is following from merlot ponte this idea that you can’t have a

7:32

concept of what a subject is without already having a concept of what an object is that the idea of

7:40

a point of view and experience infers a relationship between two different sides

7:45

a thing to be perceived but also a perceiver who is able to extract that

7:50

thing from its world and think about it as something independent so

7:56

if subjectivity is this interaction between sensation and imagination what do i actually mean by imagination i

8:02

think imagination is the ability to construct things that aren’t a mirror of the outside world so one way of thinking

8:10

about that could be novelty generation but i would define it as a cognitive process that facilitates the generation

8:16

of internal subjective representations in particular including mental images

8:23

so what is sensation sensation allows information from the world as independent cognition that

8:29

thing outside of us to implements those same subjective representations

8:37

so what’s the relationship between sensation imagination and a machine so those models sorry those definitions

8:45

of subjectivity definitely apply at least in my thinking to biological agents

8:51

but maybe the idea of machine imagination is a little less clear and so i would define machine sensation

8:57

for example as the representational information patterns that are fed to a computer and they’re

9:03

often called the training data or a corpus of information a body of information and those are most often

9:10

measurements of something it could be atmospheric measurements or an image from a live camera or almost anything

9:17

the flip side of that is imagination and i think of the imagination of a machine

9:23

as the as being implemented by unsupervised machine learning algorithms

9:29

so i’m not going to get into what that means very specifically beyond saying an unsupervised machine learning algorithm

9:35

is like a classifier it decides that in this space of measurements which

9:42

things belong to cats which things belong to dogs which things are associated with maleness which things

9:47

are associated with femaleness and i think of that classification problem as the projection of boundaries into that

9:55

underlying continuity that is independent of cognition so here we have a whole bunch of points

10:01

this is about 10 000 individuals the x-axis i believe is uh

10:07

height the y-axis is weight but i could have gotten those flipped doesn’t really matter for the purposes of what we’re

10:12

going to talk about and this is classified or labeled data so the white dots mean one thing and the

10:19

black dots mean something else so we might guess that since we’re talking about body mass that there could

10:25

be a gender thing going on here we can see that the images at the very top here

10:32

there it is above a certain point there are mostly white lines

10:38

below a certain point there’s mostly black points and then the same thing for the extremes on the other axis

10:44

and so sure enough if we plot the means which are very hard to see but

10:49

that is the mean of the black dots and that’s the mean of the white dots we can say that yes on average

10:57

men tend to be taller and heavier than women who tend to be smaller

11:02

i’m bringing this up because as we can see this data is very very messy and this idea that there are two categories

11:08

men and women as separable by some property as in body mass

11:15

means that there’s a lot of things that are lost there are a lot of women that are really really exceptional

11:23

and then there are a few men that are extremely exceptional

11:28

and so what’s important to me here is this thinking about that underlying continuity that when we’re measuring the

11:34

world less biased measurement always includes some bias but less biased by cognition

11:41

that there’s a lot going on the the classification problem the deciding which are men which and women involves

11:48

projecting these imaginary boundaries and my argument is that those boundaries are fully imaginary

11:55

because they can have multiple positions we could draw a straight line that bisects the two means

12:02

or we could invent something really bizarre and i would say a lot of the kinds of

12:08

knowledge conceptual representations that are generated by machine learning algorithms these days probably look a

12:14

lot more like this than they do a straight line so the question of arbitrariness

12:20

and kind of generative processes and imagination is always implicit the very bottom level of these systems

12:27

so i say these are imaginary because there’s no single objective solution for

12:33

dividing those two spaces that they’re conflated they’re un they are continuous in their essence in an underlying

12:39

fashion different similarity methods and different algorithms would draw different kinds of lines and that really

12:45

changes how people fit into those categories these boundaries that define groups

12:53

allow us to group sensory information into different camps so we can think of a kind of proto

12:59

concept or in the case of my work i call them percepts very low level

13:05

abstractions of sensory information that refer to a category for example male or

13:10

female and so the contents of subjectivity emerge from this interaction between

13:16

underlying data itself that there is a distribution there are measurements that can be made of something independent of

13:23

us but there’s always necessarily this act of projecting imaginary boundaries onto

13:29

that continuity and so i would say subjectivity is this emergent interaction between those two forces

13:36

between what the world has independent of us imposes upon us and what our processes of imagination impose upon

13:43

that external world independent of us so to summarize the basis of

13:48

subjectivity and also i would say that also the basis of all of cognition is

13:54

perceiving two different sensory patterns as belonging to the same imaginary class it’s seeing two leaves

14:00

and just thinking they’re a leaf and not attending to the uniqueness and individually measurable properties of

14:07

each every moment in time is a unique moment we can only make sense of

14:12

anything by grouping things by similarity so

14:18

what is the point of this uh i think it’s a useful way of

14:23

breaking down the kind of magic around a machine learning system just being a big black box of objective knowledge

14:31

because it allows us to poke it in certain ways to see what are the different components that are actually

14:38

informing the kinds of knowledge or the kinds of models that they’re actually producing

14:44

and not only that we can also use this kind of framework for thinking about both our machines and ourselves in

14:51

a way of interrogating our own bias and prejudice around the imaginary concepts

14:57

that really constitute our lived experience so we can ask very specific questions of

15:02

the black box how sensation is represented how are we measuring these points

15:09

what is preserved and what is removed how do those people who are on the wrong side of the line

15:14

fit into the model or not what are the similarity being similarity

15:20

measures being used how are we comparing apples and oranges and whether those ways of comparing

15:27

things are actually valid culturally or not and of course finally the summation of all those things is where the boundaries

15:34

between classes in this continuous space that is

15:39

beyond our comprehension where they actually end up

15:45

how much time do i have perfect i can go through this really slow and easy then

15:52

so uh watching and dreaming is my current body of work and i’m taking the kinds of

15:59

learning machines that i had been re using in previous work except i’m applying them to the appropriation of

16:05

popular cinematic depictions of artificial intelligence and right now i’m working with blade runner 1982 the

16:12

original 2001 space odyssey and tron which are all 80s movies that kind of

16:19

cue some kind of science fiction future but are also a time where artificial intelligence kind of started growing in

16:25

a really it was one of these early humps of growth of machine intelligence

16:32

so in this appropriation system the original film itself both the frames and

16:38

the sounds are the machine sensation so the world perceived by the machine is the world of

16:44

a film so the title of the work is watching and dreaming the dreaming war refers to my

16:49

phd work and i’m not going to talk about that too much and this is kind of a false dichotomy

16:54

between watching and dreaming because the argument of my entire phd is that watching and dreaming are

16:59

continuous it’s one system of simulating mental images

17:06

so watching is more an emphasis on how sensory information allows the

17:11

construction of mental images through these subjective processes and that is how we make sense of the world

17:17

that is our own kind of perceptual system but it’s also how the machine makes sense of

17:24

uh 2001 space odyssey for example dreaming is a little more complicated

17:29

but we can think about it as relating to time so dreaming is more about the occurrence of

17:35

events over time rather than the kind of construction of a

17:40

individual moment in time based on sensory information but again that’s a false dichotomy

17:46

so watching 201 space odyssey was very recently completed its first exhibition will be as part of

17:52

a triptych of works showing versions for tron and

17:57

blade runner at the surry art gallery in april and so i’m just going to step through

18:04

a concrete visual example of this process of subjectivity so this is an original

18:10

frame from 2001 space odyssey where we see hal who’s highlighted by the the light in the panel he’s a

18:17

disembodied robot with a camera that can see around and he’s playing chess which is a nice

18:23

proxy for how we used to think about machine intelligence playing games must be the alt the ultimate of machine

18:30

intelligence because it’s hard for us seeing being embodied using objects that must

18:36

be easy because they’re easy for us but it’s totally flipped in the other direction and he’s blank jazz with somebody he

18:43

kills in the film sorry spoilers

18:48

so what does it mean to project boundaries in this continuous space so we know that that original frame is unevenly

18:55

continuous meaning some parts have more color other parts don’t there’s different kinds of densities there’s all kinds of things happening

19:01

but to make things the machine makes sense of it we have to break it into pieces and that is this imposition this

19:07

imaginary projection of of lines and boundaries in that continuous

19:12

space so this is a depiction where each colored region is one piece

19:18

that is bounded by these imaginary projections

19:23

once you do that you can extract each of the pieces within those boundaries and that is

19:30

these are all of the components from that first frame and you do this or i do this for every

19:37

single frame in the entire movie which results in 47 million little image

19:44

pieces so one thing about the fact that there’s 47 million image pieces is that’s just a

19:49

massive amount of data so in my case there’s actually two layers of subjective

19:54

imaginary construction first there’s this projection of boundaries into the continuity to construct these things

20:00

then the second level is to project boundaries in the space of all of these things in order to reduce what is

20:07

perceived as redundancy so we can see some of these are all very similar to

20:12

each other so you can imagine to reduce the amount of information we could just group all those together and by grouping

20:17

them we’re constructing these percepts which are again a function of uh

20:23

making boundaries between things that are similar and things that are different

20:28

and so once you group all those together i use an algorithm that averages all of

20:34

the ones associated with each percept so all of those purple squares get average to one kind of image

20:41

all the red things will get average to something else and one component of this project is to make large scale collages

20:48

that are intended to be 48 by 48 inch light boxes so we can’t quite get a good

20:54

sense of oh no this is horribly low resolution they are about 15 000 pixels wide and

21:02

the same size as they were in the original film so this is these percepts they’re the

21:07

higher level abstractions of that sensor information but they are the 5 000 largest

21:15

and that 5000 largest tend to be the ones that are the most unusual they’re the furthest away from the others

21:22

and that’s why they have such hard edges so these really bright colors in these corners and they’re really large framed sizes

21:29

like over here are all because those are ones that weren’t easily grouped into categories

21:35

so those are those exceptional men and women that don’t fit very closely in their own classes they’re kind of out on

21:42

the edges of things and so if i use this vocabulary that’s the visual vocabulary of the system

21:49

watching is this process of constructing this mental image that is

21:54

how the machine how can we can imagine the machine sees that first frame

21:59

where each of the pieces extracted from the original are actually represented by their

22:04

nearest percept that is the higher level abstract representation

22:10

that is most similar in terms of its features for that lower level piece and you end up with something that resembles

22:16

the first frame but is also dissimilar from it and it’s somewhere between

22:22

and each of these can approach like a gaussian distribution where the edges become really soft because there might

22:28

be a lot of individual components associated with each of them

22:34

so i’m going to finish with just a little statement

22:40

by talking about machine subjectivity my hope really is that we can all consider

22:46

the degree to which both our machines and our own mental processes are biased

22:53

and not representations of reality but really

22:58

imaginary creative constructions in our minds they just happen to be synchronized by

23:04

the fact that we live in the same world so we have that same underlying sensor information and so maybe we should be considering

23:10

the validity of all concepts in terms of this idea of a subjective point of view

23:15

a subjective point of view being a particular combination of similarity measures that

23:21

is deciding things are similar are different abstract representations of things constructed from those

23:27

similarities and differences the contents of boundaries that are projected imaginarily

23:33

and also that underlying information distribution that thing that is independent of cognition the actual

23:40

numbers that informs that whole system and i think by thinking about all these things

23:46

in the context of machines and ourselves maybe we can get better at understanding

23:51

diversity and the idea of where mean fits in with

23:57

validity of statistical models and also just computational ways of knowing the world

24:03

[Applause]

24:12

uh thank you that was very interesting um next up we have sofiane

24:17

audrey is that the correct pronunciation yes okay he creates

24:23

artistic works to different forms such as robotics electronics interventions interactive installations and net art

24:31

he holds a bachelor’s degree in computer science an embassy in computer science from the university of montreal

24:37

and he’s currently as an assistant professor of new media in the school of computing and information

24:44

science in the university of maine um he’s going to talk to us a little bit about what we turn

24:50

the machine’s attention towards

24:58

christina

25:10

hi everyone thanks for being here i’m really privileged to be here with ben and allison to discuss these

25:18

questions that are just always flowing through my brain and thanks for the organizers

25:25

at art engine for the invitation so uh in my work i’m interested in hijacking

25:31

and subverting computer science and ai technologies and sort of use them to reveal some

25:38

aspects of in the end of i guess what makes us human i

25:43

most technologies are designed to be powerful to control like suzanne mentioned earlier to enslave i really

25:50

like that idea to enslave nature to enslave uh us all

25:57

uh and uh to to at the same time to maybe solve some of

26:03

our problems to cope up with our deficiencies and our imperfections but in my work i’m

26:09

interested in flipping that and revealing these very imperfections that i think makes makes

26:15

us what we are uh and i think that these uh these things can be found in these

26:21

technologies that we we design so

26:26

so most of my work revolves around these notions of agent and behavior

26:32

agents being also coming from this computer science notion of agent-based

26:39

model comes from cybernetics an agent being a an entity that acts in the world in

26:46

response to observations and the behavior being like how

26:51

we can how another entity would perceive this agent so uh uh as an

26:58

unchanging form of events due to the activity within this assembly this comes from gordon pass

27:03

pasc has a very broad notion of what an assembly he is he talks about the behavioral statute so i would personally

27:10

i i would uh argue on that but so i i

27:16

wouldn’t personally replace the word assembly with agent but i’m interested in how we experience these behaviors now

27:23

they can be used to generate new experiences so uh

27:29

this is an example of a work uh one of the first actually works where i used i started using machine learning in

27:35

2011 it is a robot that was

27:41

hanging to a cliff and the uh the catalog uh uh

27:50

uh and this robot is um looking uh at the cliff right now

27:56

because he it it it it it doesn’t like sun it doesn’t like the the the to see

28:02

the light but at the same time uh it it it needs the light because it needs to recharge the batteries with the solar

28:09

panels uh so it needs to you know once in a while once in a while look at the light and i i i gave it a a

28:17

i used a reinforcement learning uh system where uh to to let this agent try

28:23

to solve this conundrum over the summer um

28:29

so i’m gonna throw a few concepts here just to get the conversation ready but i mean like ben already uh

28:35

introduced machine learning but i’m just gonna give like uh um what i you know i i think what for me is a

28:42

a very uh good summary of uh to to try to understand what machine learning is in

28:49

comparison to artificial intelligence which is so so machine learning is a sub subfield

28:54

of artificial intelligence and artificial intelligence being this

28:59

big uh project of trying to reproduce human performance um in all kinds of

29:05

spheres and the problem is like is the following how can we get the pro how can we program machines to do things we do

29:11

when we ourselves do not know how we do these things and this is almost everything when we talk when we walk

29:17

when we make decisions most of the time we use

29:23

our intuitions we do these things without even thinking about them so if we don’t know how we do them how could

29:28

we even program them into a machine and the intuition here from a machine

29:33

at the basis a base of machine learning is well we don’t know how we do these things but what we know is that we once

29:39

did not know how to do these things and we learn how to do them by subjecting ourselves to experience

29:45

and by adapting so uh so that’s the approach machine learning uh brings is

29:50

like let’s create programs that can learn by themselves by looking at experience and observing the world and

29:56

this comes from the the first idea comes from the i would say from this 1950 cyber early

30:03

cybernetics notion of feedback so where a system makes a decision and gets

30:09

regulates itself makes small adjustments based on the results

30:15

now machine learning has again taken that a very interesting idea of this a an agent adapting to the real

30:22

world and and um once again i’m going to reuse the i really like this concept of

30:27

that that susan was talking about earlier enslavement enslavement so so machine learning has enslaved these

30:34

little agents into doing something um something um

30:39

where the the the agency of the agent doesn’t matter anymore uh what we’re after is to solve a problem so

30:47

using a model we train an agent on some data and we try to adjust it make it adjust

30:54

itself to uh to solve whatever problem we have whether it’s via it’s going to be like speech

30:59

recognition or recognizing faces or are making as much money as much money as

31:05

possible in the stock market um so i’m so this is a project i’m working

31:11

right now where i’m i’m i’m i’m following kind of this similar idea i’m

31:17

trying to unfold this uh process that uh of of of learning so

31:23

i mean like uh nowadays you if you look at machine learning algorithms we we use them every day we don’t necessarily need

31:30

to know how they work but i think that for me what’s interesting is how they work how they do this learning because i

31:37

think there are some very interesting um i i see that there’s a potential there

31:44

in looking at the system learning there’s a potential for uh revealing

31:50

again some things about herself so this is a problem a project i’m working on right now called morphosis with jose

31:55

luis mangani and jose is a visual artist really interested in

32:00

artificial life and and embodiment and so we’re creating this we’re working on this kind of ball

32:06

robot that will have a a a strange uh non-continuous body and we’ll have to

32:13

try to to learn how to grow we’re still far away from that but going

32:18

getting there slowly another project i’ve been working on last year is i started working with deep learning and

32:24

training on text so um i i’ve never read that book it’s a book by

32:31

uh sorry that’s my book but uh i i’ve yeah

32:36

i’ve never read the book wuthering heights uh because i’m french and that was not the

32:43

kind of that we we didn’t read this these kinds of classics at school we read the french classics

32:48

so anyway i thought like i trained on neural network to read it for me because it looked like an interesting book and

32:54

uh what i uh what i was interested i mean there’s a lot of like stuff like that that exists where you know you

32:59

generate systems that will you train systems on text and then they’ll be generating kind of these kind of weird uh uh

33:07

literature but what i’m interested in here is uh in unfolding again this

33:12

learning process so looking at all the little steps through which the agent goes uh

33:17

when it’s learning and this has given uh two projects that right now one of them

33:23

is an art book so it’s a series of 31 unique books

33:28

that retrace this kind of learning process and another project is uh called of the

33:36

soon this is a project with airng orange is an artist a sound art sound artist new media artist who works with

33:43

uh re-uh putting human voices in artificial bodies and artificial voices

33:48

in human bodies so we were interested in this technique uh called asmr which stands for uh autonomous sensory

33:57

meridian response which uh is you can look it up online there’s tons of videos of people doing weird weird that

34:04

are gonna the intention is that to create you know these kinds of tingles that you

34:09

you you can have when you so some some people react to different kinds of things

34:14

often it’s like very high pitched sounds like whispers um personal personal attention

34:22

you’ll see people working with all kinds of different ways to generate these states of getting

34:27

these singles and anyway uh and we thought that it would be interesting to revoice this text

34:34

generated by a an agent reading a book by a dead person

34:40

and revising it to create these this bodily try to create this bodily

34:46

phenomena so anyway um just play an excerpt and maybe we’ll see if you get the tingles

34:55

playing

35:21

is

35:55

let’s do this

36:20

is

36:32

so i don’t you could hear this kind of progress where uh so this is an example

36:37

this is really from the beginning of the composition of that text where initially the agent is completely random

36:44

so it doesn’t really know anything and then as it reads as it proceeds to read the book it learns some

36:52

rules about combining elements so of course at first it gets crazy about spaces because it

36:58

just sees that there’s a lot of spaces in the in that book uh and then eventually it becomes more

37:04

and more structured and starts creating more themes that are longer and eventually it’s able to compose

37:10

words and full sentences

37:19

so so another uh which brings me to this next aspect that

37:25

i’m interested in which is i think is relevant to this discussion is the question of inter indeterminacy autonomy

37:32

and command so uh so like uh

37:37

one of the work uh that chris alter was mentioning earlier uh that i’ve worked on uh where i i started to

37:45

uh work with these questions of um of using machine learning for

37:51

art making for generating different kinds of behaviors that would uh

37:57

feel like feel like feel like something recognizable but yet

38:04

always lying beyond beyond our grasp and and polytope we uh were investigating

38:09

trying to start with the question of um um uh how would janice zinicas

38:17

in the in 1967 and then in the in the 1980s series of work called the polytopes with these kind of like

38:24

movements of light and sound and these kind of total spectacles uh

38:30

uh uh and and we were we started from this question what would what would zenecus

38:36

how would zinecus have worked uh how would he have made such works if he had

38:42

access to the kind of technologies we had today um this is another work uh vessels

38:48

in collaboration with stephen kelly and samuel central bank where we uh create this we created these

38:54

small groups of robots that are autonomous and that just

39:00

communicate with one another and move on beds of water we’ve we’ve shown that um

39:06

uh outside in the outdoors in like public um spaces and

39:12

uh and the yeah like these robots are sort of evolving together a kind of

39:17

group behavior uh and uh and how could we use that as a

39:22

way to so so to to to um to

39:27

induce different kinds of uh ways that people are are projecting themselves on the robots and

39:34

um and so so when when i look at this work with what i find the most interesting is to look at

39:40

how people uh interact with one another uh when they

39:45

when they’re surround they’re looking at the robots and moving around the the space

39:51

um which leads me to the second principle of machine learning as we know it today uh

39:57

uh which is self-organization the idea that like you have all these independent units uh

40:03

that are uh working together to create these uh

40:08

layers of emergent representations

40:15

this is an example that i really like to i think is really telling in the 19 mid 1990s

40:22

adrian thompson i think his name was was working on evolvable uh circus so he

40:27

designed this system with uh these circuits that can they were called uh

40:33

programmable field arrays or something like fgpas and basically these are so uh

40:38

so he he tried to uh solve a very simple problem where in the input you would

40:44

send the sound wave and in the output there should be a signal let’s say you know i don’t know five

40:50

volts if the input signals is one kilohertz and zero volts if it’s ten

40:55

kilohertz something like that so just like make a classifier using a circuit and uh he used a genetic algorithm

41:02

approach where he he he tried the different kinds of circuits took the

41:07

best ones and then made modifications on them sort of have the have them

41:12

uh mix them together and then go to the next step so it is a iterative machine learning sort of approach and and in the

41:20

end he uh he got uh you know a circuit and what he did is that like what is traditionally done i guess in that field

41:26

is that he removed all the parts that did not matter uh that were disconnected from the rest

41:33

and after removing them he realized it didn’t work anymore and then he realized that actually some

41:39

of the parts in the the the the airway harry needed to stay there

41:45

even if they were not logically connected to the rest because the circus somehow had

41:52

was using them was using the small little interferences

41:58

in the circuit to achieve its goal so i think that to me this is a very interesting example because it it what

42:05

it tells me what i think we can get from that is that uh adaptive behaviors or or machine

42:12

learning systems um it’s kind of a metaphor i guess that we cannot really understand them rationally they’re like

42:21

uh it’s related to what ben was saying about deep learning we can’t really uh

42:26

they’re hard to uh to analyze we can like really uh it’s very difficult to

42:32

because they’re not rule-based they’re they’re not based on uh specific rules they’re based on uh

42:38

an adaptation of a body and i think that what i speculate is that then if we

42:45

generate behaviors with learning machines uh we cannot you you and i’ve i’ve

42:52

with the experience i have with audiences there that are looking at these these systems they’re often like

42:58

if these if these if they stay long enough so if they don’t stay long

43:04

enough they usually don’t find it interesting but if they stay long enough they get to know them and then

43:11

then they’re able to start making inferences and then it becomes more uh telling for for them so because

43:18

and i think that that that the reason is that again because they cannot be explained by saying oh this is what’s

43:25

happening you know ax like a like a causes bq which causes c uh

43:30

they can only be known uh in a phenomenological manner

43:36

um so so how do we deal with inter indeterminacy uh common versus uh

43:43

autonomy uh so uh i’m just just saying like i think this is a thing

43:49

that is coming very often in and when you work with such emergence or sufficientizing machine learning systems

43:56

is the tension between comment and autonomy so as an artist you will have certain intentions you want to

44:02

make something so you want to keep control over things and this is a bit with the zenikus voices comes and within

44:09

a case he was using these indeterminate systems and he was saying you know i’m the ultimate arbiter of these processes

44:16

and because kind of the opposite view and i’m i guess i’m closer to that other view which is comes from uh john cage

44:24

who says well let’s sound be let sounds be themselves that let things be themselves so let’s just give put some

44:30

things out there leave these things to be autonomous it might be good it might be bad i don’t

44:36

care what’s important is to there’s there needs to be this um the the

44:42

these autonomous processes have the potential to uh bring certain certain

44:47

degrees of surprises uh uncanny feelings that i’m interested in

44:54

uh and i will just finish with that uh just talking a bit about representation which

45:00

is the third principle of deep learning so deep learning starts from

45:05

these uh feedback system connectionist self-organizing systems

45:11

but adds another level by saying like uh but by actually finding ways to

45:17

to to move with through different layers of representation from very simple

45:24

uh so if you train a neural network on faces for example

45:29

i think this is kind of a example of a convolutional neural net in the lower lower lower layers you’ll find edges

45:37

like this automatically found like very simple constructs and as we move uh

45:43

higher and higher in the arrow keys you get more and more um

45:50

higher level abstractions that get reinforced

45:55

so uh i guess i’m am i out of time yeah okay so i will

46:01

just show very quickly some crazy images to get your attention just

46:06

very quickly this is a new project i’m working on where i uh with tez

46:13

and we are trying to get uh to create this loop of uh this feedback loop between

46:22

uh living organisms called ooglin aggressively which are kind of photosynthetic

46:28

microorganism that can move towards the light and you can see that to the naked eye you can if you project light on them

46:34

they will assemble in a way that they that that i think evokes a certain form of

46:41

self-organizing um perception and then these images are sent into an alkaline

46:47

encoder which is a kind of neural network that we twist in a way that like this

46:52

this auto encoder is going to try to it’s going to look at an image and try to imagine

46:59

an another image that would look like a like a digit

47:04

like from 0 to 9 but but a handwritten digit but because the input is maybe not

47:11

is noisy it will actually come up with something strange and and new and this is an

47:16

example here in the corner on the left and up corner

47:23

and this is another so kind of weird glyphs that are not really digits but are

47:29

maybe examples of the kind of patterns imagined by this uh collaborative system between

47:36

microorganisms neural networks and humans who are actually at the source of the database

47:43

it was used to train on so

47:48

thank you

47:54

[Applause] uh thank you very much um so last up is our speaker

48:01

allison parish she’s a computer programmer a poet

48:06

educator and game designer who’s teaching and practices addresses the unusual phenomena that

48:12

blossom when language and computers meet with a focus on artificial intelligence and computational creativity she’s a

48:18

teacher at nyu interactive telecommunications program where she earned her master’s degree in 2008.

48:24

she was also named best maker of poetry bots by the village voice in 2016 and

48:29

she’s been published her her computer generated poetry it’s been published in ninth letter and veg

48:37

and uh you’re going to talk to us today about computer generated poetry

48:42

something like that um

48:51

okay so i have hi everyone i have my notes on my phone and i realized only when i was sitting

48:58

up here that i don’t have enough hands to operate all of these things at once so

49:03

let me see if i can that’s that’s almost good right maybe if

49:08

i turn that upside down you can signal me to click nope i can do

49:15

it all right okay i can i can successfully move

49:20

forward in my slides and backwards look at that i’m like a cyborg um

49:27

so um hello everyone um i’m allison parrish i’m a poetic programmer and game designer um i

49:34

am a professor at the interactive telecommunications program at new york university

49:39

um where i teach a class called reading and writing electronic text which is a class about computer generated poetry

49:45

and learning how to program in python my most recent

49:53

project that sort of illustrates the kind of work that i do

49:59

is this book of poetry that was just released by counterpath

50:04

just last month it’s called articulations and i included a sample of the output here

50:10

this book uh started out as an exploration of the idea of similarity and poetry i’m actually hearing ben

50:16

discuss um discuss this project i realize that there’s actually a lot of similarities between this project and his

50:22

um so i started out with the idea of similarity in poetry and how similarity whether phonetic syntactic or semantic

50:29

contributes to our sense of poetic-ness um and cohesion in a text the poetry is

50:36

constructed from a corpus of several million lines of poetry that i extracted from project gutenberg which is a

50:41

database of thousands of texts that are in the public domain

50:46

and then i used a phonetic dictionary called the cmu pronouncing dictionary to create a computational model of

50:52

phonetic similarity that predicted a vector or like a sequence of numbers like a point in space

50:58

for each line of poetry then those vectors have the property where um two lines that sounded similar to each

51:05

other are closer the vectors are are more similar if the lines of poetry sound similar to each other um

51:13

so the composition itself is accomplished by doing a random walk through that space essentially it starts

51:19

with the randomly selected line of poetry and then finds the line of poetry that is most uh phonetically similar to

51:25

that and then adds that to the output and then finds the line of poetry that’s most similar to that and adds it to the

51:30

output um and so forth until it’s produced like you know a book of of

51:35

poetry and it excludes any line that’s previously been in there um the end result is a kind of poetry where

51:42

cohesion is achieved only through phonetic similarity that’s the only organizing principle of this text is

51:47

that the lines sound similar to each other um it’s weird and it’s a lot of fun to read so i thought i thought i might um since

51:55

i am a poet i should read some poetry the other the other artists here today have the

52:00

benefit of being able to show videos of their work that sort of like you know express what they’re about but i feel

52:06

like just seeing this on the screen you don’t really understand what it’s like so i’m gonna i’m gonna read it um

52:13

sweet hour of prayer sweet hour of prayer it was the hour of prayers in the hour of parting hour of parting hour of

52:20

meeting hour of parting this with power avenging his towering wings his power enhancing

52:26

in his power his power thus the blithe powers about the flowers chirp about the flowers of power

52:32

a butterfly must be with a purple flower might be the purple flowers it bore the petals of her purple flowers where

52:38

the purple aster flowered here’s the purple aster of the purple aster there’s lives

52:44

a purpose stern a sterner purpose fills turns up so pert and funny of motor

52:49

trucks and vans and after kissed a stone an ode after easter an iron laughter

52:55

stirred a wanderer turn a wanderer return a wanderer stay a wanderer near

53:01

ben a wanderer i wander away and then i wander away and then we shall we wander

53:07

away and then we would wander away away oh why and for what are we waiting oh

53:12

why and for what are we waiting why then and for what are we waiting a little excerpt of

53:18

how that feels thank you [Applause] so

53:24

articulations is a part of a series of computer

53:29

generated books um released by counterpath um whoops let me come back

53:35

this is a mess um released by counterpath and the the series is called using electricity and it’s named after a line in allison

53:42

knowles house of dust which is a early important work of computer poetry

53:47

um the series is edited by nick montford who also showed up in in sofiane slides and nick’s book was also released as a

53:55

book in this series that’s nick’s books there it’s called the the true list

54:01

these books just came out and we did a little tour of the northeast united states together

54:06

along with rafael perez perez who was also whose book was also released in the series recently

54:11

um nick’s book is also a book of computer generated poetry um in in this case it

54:18

uses a small data set and a deterministic algorithm to generate a sequence of english compound words like um waterfield moon

54:26

hill ring sac windman i urge you to seek out this book and recordings of nick reading it because

54:32

it’s really um it’s really remarkable and entertaining and it makes your brain do interesting things

54:38

on tour when he’s introducing this book he would talk about the word true because the book is called the true list

54:45

and he points out that this word means not necessarily factual but the opposite

54:50

are not necessarily the factual in the sense of the opposite of false but in the sense of

54:56

conforming to a pattern or in the sense of the verb to true which means to straighten or to align or to make even

55:02

with something like when you true a painting that’s hanging on the wall or you true the spokes of a bicycle wheel

55:09

um this is gonna get weird

55:15

i might have to um it’s feeling good so far this weird

55:21

thing that i’m doing maybe i’ll try it this way now um no that’s not gonna work here thank you

55:28

um no we’re good we’re on the right side now okay um so simultaneously i’ve been

55:36

reading i’ve been rereading the book ted berrigan’s ted bergen’s book the sonnets which is

55:42

this really amazing book of poems composed with a cut up process similar to the one used by burroughs and

55:49

brian guyson um this is maybe my favorite book of poetry of all time i should have included an excerpt over

55:55

here um but um the that book is from 1964 berrigan was

56:00

sort of working in the same era as cage and the fluxus and fluxus and all of the

56:07

stuff happening in that scene and this is a quote from the introduction to the book um alice notley

56:12

wrote this introduction and i’m just i’m just going to read this for cage the application of a method

56:18

results in a work or a performance that’s serene and free or rather liberated as cage is a serious ben zen

56:24

buddhist and his work is permeated with buddhist thought for ted it more results in one that’s

56:31

deeply true an alliatory method used with seriousness and respect puts one in

56:36

touch with hidden powers and truths within oneself and becomes revelatory hopefully in a way that might be

56:41

relevant to anyone yet by giving oneself up to chance an artist does not lose his or her originality as marcel duchamp

56:48

said um your chance is not the same as mine just as your throw of the dice will rarely be the same as mine

56:55

and that was really interesting to me uh next slide please

57:00

um so i sort of i take it as axiomatic that

57:06

cut up procedures are a kind of computational procedure even if ted barrigan wasn’t working with a computer

57:12

it’s still essentially a computational process to take a text and cut it up and randomly arrange its components or or

57:19

other techniques for rearranging a text um but that formulation that that it was

57:25

enabling ted berrigan to do something that was more true that really struck me um

57:33

and it sort of helped me put this question to to my research can computers help us write poetry that is true and

57:40

true again in that sense not of like being the opposite of false being factual but

57:45

instead poetry that is that is firm and that’s fundamental um that expresses

57:51

something um that that is true right

57:56

um next slide please um of course most people don’t think of

58:02

computer generated poetry in these terms right um usually the idea of

58:07

computer-generated poetry is framed like this um where that that narrative of the labor of

58:13

artificial intelligence where any any ai task is essentially just trying to take over the labor of a person um and

58:20

the idea that you know human expression will be superseded by cold robotic precision eventually and we just have to

58:26

go along with it we’re helpless to resist right um and that relies on this idea that you

58:31

can formulate the composition of poetry as a task that boils down to just like a list of instructions which you can with

58:38

research be reproduced with perfect fidelity and if you look at a lot of the like a lot

58:43

of the academic research is happening in in artificial intelligence and computational creativity right now it’s

58:49

all based around this idea of how can we specifically create poetry that tricks people into thinking that it

58:55

was written by by human beings in a typical way like even reifying the difference between human and machine

59:02

authorship which i don’t actually think exists and hopefully you’ll see that by the time i’m finished speaking

59:08

um but that idea that the the purpose of using artificial intelligence and art is

59:14

to recreate with as much fidelity as possible things that are produced by humans without aid of those techniques i think

59:20

is like extremely boring like it’s the worst possible thing to do with this technology

59:25

um next slide please um when i’m making poetry with a

59:30

computer i’m actually thinking about this line of argumentation that was you know it’s been articulated by many people in many

59:37

different ways um but viktor shklovsky in 1916 and in that

59:43

wonderful old saw which i can get to somehow in my notes

59:49

artist device he says the purpose of art is to impart a sensation impart sensation to an

59:54

object is something seen rather than merely recognize the divisive art is the

1:00:00

device of estrangement of things and the device of defacilitated form enhancing the difficulty and duration of

1:00:05

perception so the perceptual process in art is an end in itself and should be prolonged give me the next one there ben

1:00:12

thank you um and then talking specifically about poetry he says in studying poetic speech

1:00:19

and its phonetic and lexical structure as well as in the characteristic thought structures compounded from the words we

1:00:24

find material obviously created to remove the atomicism of perception right so in other words the the goal of poetry

1:00:31

is to make language unfamiliar by making language unexpected um next slide please

1:00:38

um so tristan zara who has uh taught us uh

1:00:44

writing about the same time as uh schlovsky actually um had like sort of like the the

1:00:50

prototypical idea of how to um make a language make language unfamiliar this i think is like one of

1:00:57

the earliest examples of a computer-generated poem even though again it’s not generated by a computer it’s just a procedure that is sort of

1:01:04

computational in nature and in this technique the poem is composed not by arranging words um by

1:01:10

intention in the way that we usually associate with poetry especially like the romantic poets like you know

1:01:16

wordsworth or whatever um instead the poem is composed from a source material that’s taken from the surrounding environment like a newspaper in this

1:01:22

case and then arranged according to a procedure and that procedure is um to select a word at random and add it

1:01:29

to the end of the poem and then remove that from the pool of potential words and with this procedure you’re very

1:01:34

likely to end up with an arrangement of arrangement of words that’s never before been seen in recorded history right um

1:01:43

and it’s nonsense but it’s nonsense that is a result of a process that can be understood

1:01:48

right um now there’s one line from this

1:01:54

that i think is important um we can go over like what exactly he says in here um take a newspaper take a

1:02:00

pair of scissors choose an article as long as the poem that you want to make cut out the article got out each of the words blah blah blah blah copy

1:02:07

conscientiously the last point there i think is this like you know very typical like uh masculine uh sarcastic

1:02:14

distancing from the amazing thing that you just did like oh i invented this process and it’s really cool but now i

1:02:20

have to like stand back from it and say that i’m actually better than it um but it’s that line right before that

1:02:26

that i think is most interesting um can you give me the next slide there please ben

1:02:34

the poem will resemble you right um

1:02:39

and note that zara didn’t say the poem will resemble the reader right he’s not just talking about

1:02:45

apophenia here he’s not just like talking about the idea of picking out patterns and randomness

1:02:51

um he’s talking to you the writer like it was how to write a daughter’s poem he’s

1:02:57

talking to the person who’s following that that tutorial right um the poem made with this procedural

1:03:02

remember will resemble you the writer and i keep thinking about this line and um

1:03:08

for me it succinctly and precious presciently addresses algorithmic bias

1:03:13

like the idea that if you take the world and cut it up and decide how that cut up process will proceed

1:03:19

and how the world will fit back together it’s about it’s about you

1:03:24

and i talked to nick mumford about this and he um i don’t know we we can all talk

1:03:30

about this here i don’t know i don’t know if these these two um would also think that their work resembles them

1:03:36

that it’s saying something about them as individuals but i think it’s it’s very clear that it is like all of these decisions that all of us are making with

1:03:43

these techniques are about our own aesthetics at some at some point and about our own lives

1:03:48

um uh next slide there

1:03:54

oh you’re already there oh wow we’re having some some machine learning happening up here right

1:04:00

distributed cognition yeah um so um

1:04:05

this is uh jackson macleo a quote from jackson mcleod who is also a contemporary of cage and who also worked

1:04:11

with um techniques uh using randomness and chance and procedure

1:04:17

um and i think this this quote sort of sums up how like the sort of the mental

1:04:22

progression even for for me in making the kind of work that i make and how it moves from this idea of like

1:04:29

well what i’m trying to do is i’m just trying to like un expose like the under like the underlying structure of something that i’m trying to um

1:04:36

i’m trying to uh remove the ego of the artist from the process right but um

1:04:41

jackson macklev says the motive for the use of chance uh means etc uh was to be able to generate a series

1:04:48

of dharmas relatively uncontaminated by the composer’s ego it was such a relief to stop making artworks that carry the

1:04:54

burden of expression but he says i do allow my own emotions to influence my systematically generated work my choices

1:05:00

of means materials etc can’t help being influenced by emotions and i’d be foolish if i thought they weren’t

1:05:07

right um and from this what i’m what i’m realizing and what i take away and what

1:05:13

i’ve been um concentrating on almost in my own work is that computer generated poetry isn’t

1:05:19

about automating the work of the poet but instead amplifying that work and

1:05:24

letting it accomplish things would other be otherwise be difficult or impossible to accomplish

1:05:33

no that was good i was talking about the idea of nonsense

1:05:38

before and um nonsense is like usually something that has like a negative connotation

1:05:44

like we don’t like nonsense it’s something that that we would prefer to avoid but nonsense at some point is like

1:05:50

sort of the point of poetry it’s like we’re trying to make language that’s unexpected language that has never been

1:05:55

heard before people resist it but um poetry is about making language unfamiliar and language that’s

1:06:02

unfamiliar is difficult and frightening and sort of the role of the poet in my

1:06:07

view is to make nonsense that that follows systems that that people can be led into by the hand so that they can

1:06:14

undergo these processes that use their sense of how to understand language in ways that open up their minds to to

1:06:21

different ways for reality to work um

1:06:26

i think this is like yeah this is the last slide so i can’t i’m just eyeing you out okay all right i

1:06:32

can’t tell if she’s being impatient with me or not um so i haven’t like focused a lot on like the

1:06:38

specifics of any machine learning stuff in this uh talk that’s because i i for

1:06:44

me i don’t really see artificial intelligence and machine learning as being categorically different

1:06:51

from the perspective of the artist from techniques that just have to do

1:06:56

with chance or procedure like to me it seems to stem directly from that and a

1:07:01

lot of the same artistic moves and you know rhetorical moves that you can make with those other previously used and

1:07:09

still frequently used types of work types of algorithmic work are very

1:07:15

similar to the ones that could be done with machine learning specifically um but the things that you can do with a

1:07:20

computer that are that tristanzara couldn’t do with a pair of scissors in a newspaper

1:07:26

have to do with um data and scale like i can make work that

1:07:31

um whereas if tristan zara wanted to make a cut up of like a newspaper that was fine

1:07:36

um if you wanted to make a cut up of every newspaper he wouldn’t be able to do that because of the labor involved right um but i as a poet can make a cut

1:07:44

up of like every newspaper i can find a corpus of like you know every newspaper every pub ever published or something

1:07:50

close to that and then use that cut up technique on that and then see how that’s different like what’s added to it

1:07:56

by that um the other difference is scale which goes along with that like i can do things um very quickly with a lot of

1:08:02

data um and those are the things i think that really open up the possibilities for artists working with these

1:08:08

potentials and machine learning is really just a way of like sort of managing that that scale um

1:08:17

yeah so when i’m talking about my own work i see it you know directly as um as

1:08:22

an extension of the kind of work that jackson macleo and justin zara and ted berrigan were doing with chance

1:08:28

procedures um coming up with new ways to compose language using procedure and and chance

1:08:34

um and then i have my my social media info on that side so

1:08:40

um that’s it [Applause]

1:08:50

okay um thank you all for three really very interesting presentations sorry i’ve

1:08:56

been taking notes while you’ve been talking i’m not just like on facebook or on twitter um

1:09:02

yeah so it’s interesting though to see as well just there’s a lot of correlation between all of your practices in terms

1:09:08

of like how you’re thinking about procedures and processes and the types of things that actually machines actually open up

1:09:15

um one of the other things i was really quite interested to actually hear about you as well is that you’re all kind of

1:09:20

interested in the kind of like the perspective of the audience and how you can kind of bring the audience in

1:09:26

through these processes as well so i have kind of two questions that relate to the perspective of audiences um the

1:09:33

first question i wanted to ask which is a little bit about kind of how

1:09:39

we imagine ai as an audience or the public perception around it so for me a

1:09:44

lot of the idea of the applications of artificial intelligence as a broad like

1:09:49

is it as a field as a domain but right now the way that it’s used and the way that it operates it operates as a kind

1:09:55

of as a tool of capital it’s there to for monetization it’s there to improve

1:10:01

processes it’s there to reduce labor so it has very much kind of

1:10:06

its foot kind of firmly planted into the field of um like technological growth and kind of a

1:10:13

capital growth so but when it comes then when we look then kind of like how do we understand kind

1:10:19

of like ai as kind of uh together and that’s kind of where maybe the

1:10:26

question around audience comes into play so most of the narrative that we as kind of like when you think about people who

1:10:32

are not really and who don’t kind of like inca or think about kind of artificial intelligence in a very kind

1:10:38

of day-to-day sense a lot of the way that or a lot of the narrative about advertisement intelligence comes from

1:10:44

science fiction and literature and so you know a lot of these things they speak to this idea of like sentient

1:10:50

sentient machines or utopias and dystopias and things like that and really my question here is what

1:10:56

where is your practice located in relation to either one of those things in the real world capitalist ai

1:11:03

or in the sci-fi fantasy ai and it’s also perfectly okay to say neither

1:11:12

i mean i can’t get away from ai in cinema because i’m using ai and

1:11:18

cinema as a reference point uh so it’s hard for me to answer that one

1:11:24

because i mean i guess i would put it in the sci-fi category in general

1:11:31

but at the same time as a kind of antidote to the sci-fi

1:11:36

positivism while i mean just the sheer fact that

1:11:42

i’m appropriating copyrighted material and repurposing it in this weird artistic context

1:11:47

is a rejection of some of the axioms of of capitalism because i’m you know

1:11:53

appropriating the labor of others in some domains illegally

1:11:59

that’s arguable but that’s definitely part of the work you know you can’t get away from the fact that i’m yeah

1:12:05

uh taking some you know creative work of others but creative work in a context of

1:12:11

uh value generation in terms of capitalism so i would say both neither and maybe a

1:12:18

little more of the ai one i mean the science fiction one

1:12:27

um i would say that

1:12:33

it’s complicated uh part of the reason why it’s complicated

1:12:39

for me like uh there is i and i guess it’s true for youtube but i i have a

1:12:46

all of the things i’m creating with ai machine learning comes from my own

1:12:52

experience going uh through uh my own experience of working with these

1:12:57

systems and i i i i i i my first uh

1:13:05

my first goal in life was to be retired like my grandfather but my other uh alternative job was to be become an

1:13:12

inventor and become someone who would become a scientist computer scientist working in ai and i i had actually the

1:13:19

chance to study uh in the uh early 2000s with in yosho

1:13:24

banjo’s lab at the university of university of montreal so i studied the neural networks before

1:13:30

they became uh marketable and then i i left to become an artist

1:13:37

and i never regretted that choice and and uh and and i i think that part of what i’m

1:13:45

trying to do in with in my work is i’m always sort of trying to

1:13:50

work with these um these capitalistic technologies and just like

1:13:55

just turn them into something that’s uh that does not have any market value

1:14:02

that’s that’s uh that and something that interests me and so the reason why i

1:14:08

first got into this field was that i’ve been like uh i was i’ve been obviously fascinated at first

1:14:15

like a lot of people go into ai by science science fiction and um

1:14:21

but uh but i’ve always been uh

1:14:26

just viscerally uh attached to seeing these uh yeah these these these things running in

1:14:33

the on on the machine that can learn stuff that can like you know become

1:14:40

that can transform and and and and and just through their own perception uh

1:14:46

learn something about the world so i guess that like i’m not i’m not nowadays

1:14:51

i’m not thinking about making anything that is science science fiction i’m just thinking about

1:14:57

bringing uh bringing to bringing to people who are

1:15:04

um maybe uh just breaking people’s uh

1:15:12

biases about these technologies so uh

1:15:17

by like for example when i show the the robot installation

1:15:23

we have people who are passersby and they think that

1:15:28

you know we’re trying to like they think that we’re we’re there to do an advertisement uh they’re

1:15:34

they’re that we’re they’re afraid i’ve seen people attacking the robots trying to yeah i had to jump a few times to

1:15:41

save the robots and the water and stuff like that or or then and then some people become you know

1:15:48

that there’s i create i guess i create the try to create a space

1:15:54

for people to imagine something alternative about uh

1:16:02

about these systems um so i guess this is so so in a way it’s like

1:16:07

yeah it is definitely anti-capitalistic it is i’m not seeking uh

1:16:14

to necessarily to um yeah to to to to have people think about

1:16:21

i’m not necessarily thinking about science fiction i’m thinking about yeah what what are like what are other things

1:16:26

we can do uh with these technologies and how can we how can the audience be

1:16:32

participate to that and and and get to maybe develop a different kind of

1:16:38

feeling about technologies that there is not something that’s separated from us

1:16:43

this is just part of uh our of society that this is not something that necessary needs to be

1:16:50

imposed by other power structures that you it can actually be become just part of that it

1:16:56

can resemble you that it can the technology can fail that technology can be imp imperfect uh that it can be cute

1:17:04

and curious and and and uh and and frustrating and that this

1:17:11

is something that is you know just and yam just helps us maybe articulate how we

1:17:17

how we live together so anyway yeah yeah i um i

1:17:24

i sort of feel the same way that sophian does about that i think um the

1:17:30

the underlying critique that that i try to make apparent in my

1:17:36

work is just like by using computational processes and and

1:17:41

machine learning and artificial intelligence stuff um people can come to understand because i

1:17:47

do try to make work that like shows the process that that produced it on the surface as much

1:17:53

as possible because i think that’s what makes both good art and good poetry um

1:17:58

in doing that i’m hoping to like you know help people to better understand like oh this is what the computer is

1:18:04

capable of and also that this is this is all the computer’s capable of right like

1:18:10

this is the extent of it and this is also like um if an artist can do it then i could do

1:18:16

it too like it doesn’t go because artists as we know are like the least capable of the professions

1:18:24

um but yeah just just like trying to show through um through the work like

1:18:30

you know here’s here’s how these underlying systems work um i think that

1:18:37

one of the one of the jobs of poetry as a medium is to show like um that language could be otherwise that

1:18:44

we could we could speak differently and use language in ways that are that are unexpected and have their own systems of

1:18:50

meaning and to the extent that language shapes reality and shapes politics that’s a very um that’s a very

1:18:56

political message for poetry to be able to make and so that’s sort of what i’m trying to

1:19:02

do with uh poetry that uses machine learning is to say like you know talking could be otherwise here here are

1:19:08

all of the beautiful ways that we can make talking otherwise yeah yeah i mean i guess it’s one of the

1:19:15

things i really like that sophian just said in his talk was just about like how do we program machines and we don’t really

1:19:21

know how to when we don’t even know how to do things ourselves i think that’s kind of like quite inherent in terms of

1:19:27

like the model of society that we live with that is kind of a little bit messy and

1:19:32

that doesn’t really work for everybody so it’s in h it’s definitely kind of like apparent that like true

1:19:38

calling attention to true procedures and through the processes by which kind of machines make decisions that maybe it

1:19:44

opens up a space of understanding that is not necessarily embedded in this kind of like science fiction space

1:19:50

of like oh ai is and it gets people to see it more as a tool rather than a

1:19:55

piece of kind of like fantasy or magic i think that’s a really important thing especially when it’s a very powerful

1:20:02

piece of knowledge to give to an audience because especially when you think about kind of like how technology

1:20:08

is designed so like this morning i don’t know if anybody saw but like today google

1:20:13

removed the view image um if button

1:20:19

right click um so that now when you go on to google searches and you look like

1:20:24

view by image you can no longer like uh view the image or like copy and paste it

1:20:29

and they do it as a very kind of deliberate strategy to kind of control kind of information or to control kind

1:20:36

of the circulation of information in a very particular type of way as well so i guess the second question before i

1:20:42

kind of like open it up to the audience a little bit is to maybe kind of think about oh where did i do my notes

1:20:48

um i guess it’s just that it’s a it’s a bit of a broad question

1:20:54

and i guess one of the things that i’m kind of interested in kind of knowing understanding from your perspectives is

1:21:00

like what does ai is a tool offer for the interpretation of artworks

1:21:06

and how do you think that maybe it progresses our understanding of of kind of like the way that that art could

1:21:12

potentially be as well because you’re all using art in a very like in a creative

1:21:18

um creative domain and i really like this idea of like creating stuff that’s useless i think that more artists should

1:21:24

definitely do things that are useless rather than having to be make something that has certain sort of like value or

1:21:30

it becomes this kind of like thing of pointy art which is basically just pointing things out with your art

1:21:36

so i guess yeah that’s my question is like how do you think audiences can be drawn into kind of new understandings of like

1:21:44

where art can go i guess is that too fake

1:21:50

well i have a thought just before which was also that we can use science fiction also as a tool to reflect on ourselves

1:21:57

and just to say it’s this hypey future thing i think is underestimating it a

1:22:02

little bit i think there is potential for thinking through science fiction as a way of reflecting on ourselves and so it

1:22:10

i think of and that’s the reason why i chose science fiction films is not just the depiction of ai but my machine is a

1:22:15

way of thinking through our own processes swallowing this other cultural machine

1:22:21

of us understanding our own processes that is science fiction uh in terms of the next one i

1:22:28

i’m conceptualizing another project at this point which is uh kind of taking the idea of the art

1:22:34

market and twisting it in a way by making a very banal

1:22:40

aesthetic generator that will just do whatever people pay attention to and is intentionally not very

1:22:47

interesting like machine zombie formalism exactly

1:22:52

it’s called the zombie formalist good good catch you got the reference there so it’s just weird so that it is

1:22:59

intentionally unvaluable but at the same time making fun of value

1:23:04

and being it’s supposed to be ironic and a lot of my work is actually a little bit ironic like the idea of making

1:23:10

machine dream was supposed to be ironic but things exciting

1:23:15

well it just it yeah all of a sudden things changed and the idea of a dream machine isn’t so

1:23:21

weird as it was when i started in 2009 so irony is something in there and i think

1:23:29

that kind of relates to your question is of use because you can use something and use it in a weird context in an ironic

1:23:36

sense where you’re not really using it seriously you’re just using it as a way of talking about some broader structure

1:23:42

and that’s how i’m using it in that context um one one way that

1:23:51

i’m i’m trying to to name the thing that i’m thinking about i think a lot about um

1:23:59

there’s there’s a school of poets the conceptual poets um who are sort of like led by by kenneth

1:24:05

goldsmith um who’s like you know well known for making uh works of poetry that are like

1:24:11

sort of about um appropriation and scale like taking

1:24:17

somebody else’s work and taking a lot of somebody else’s work and then presenting it it’s like you know here’s

1:24:22

a total transcription of one issue of the new york times or something like that and he talks about and he’s written and

1:24:29

spoken about this he talks about the idea of um him not wanting to have a readership he wants to have a thinkership in other

1:24:36

words that the work exists purely to be thought about as like this is an artifact that i made

1:24:43

that it’s literary it consists of words but it’s not meant to be read um

1:24:48

i’m i’m the opposite of that and i think i think that a lot of a lot of work that’s that’s

1:24:54

computational or focused around procedure and especially working machine learning tends to be about that

1:25:00

overwhelmingness of scale and that seems like a message that it’s sort of like stuck on that’s like oh there’s a lot of

1:25:06

data look at just tons of data so much data oh my god what are we gonna do with all this data um

1:25:13

whereas what i’m what i’m trying to do is like i i put a lot of effort into my poetry even though it’s unusual to make

1:25:19

it something that is that’s readable and and intimate and and

1:25:24

tries to like engage you from one moment to the next um and i think that’s for me sort of the

1:25:31

antidote to that is like you know machine learning does take a lot of data but what if you’re what if

1:25:37

we’re using that to make works that you feel like they’re smaller and more intimate and are are more expressive

1:25:43

or something like that of course it’s i’m not always successful at making something that’s that’s engaging on a moment-to-moment basis but

1:25:50

that is the the goal

1:25:56

um so in relationship to your to the to that

1:26:01

question i i i will bring up a kind of a

1:26:06

radical shocking uh idea is that i i uh alison was talking earlier about

1:26:13

computational creativity which is this subfield of ai that tries to solve the

1:26:18

problem of creativity um you know by by

1:26:24

by basically having uh trying to pass the turing test and certain

1:26:30

artistic domains um and so so what i think is that they will

1:26:36

be successful and to a certain extent they will be successful in that uh

1:26:43

they will be successful in trying to solve the turing test um but for

1:26:48

um for what i would say is the formal aspects of art so

1:26:55

i mean the first science of this is in in in music you already have companies that

1:27:03

are uh now using machine learning to you know uh you don’t have to if you

1:27:09

want to have music composed for your commercial or your short movie or for your show

1:27:14

whatever you can just hire this company and they’ll have a a program uh

1:27:20

help you compose your scores so uh and there’s been like 20 years of research

1:27:26

at least done in that field so and it’s it’s ripe now for the taking now this being said this will come up with um

1:27:34

um with yeah with with uh um

1:27:40

um yeah with with the formal aspects of of music composition within a certain

1:27:47

uh area of different genres uh and uh i think you’ll see that then uh

1:27:53

coming to like you know all of the pop music uh being just

1:27:58

created entirely by algorithms it may be performed by iterasol artificial uh by

1:28:05

robots or like as some of you might have heard of atsunumiku and other vocaloids so like holograms that are

1:28:12

performing and people go and see their shows and now it’s like the the last missing step is just that the whole

1:28:19

process is being i think like i don’t see why i i think this is a low-ending fruit and

1:28:25

it’s difficult but you know i think that it’s given given the the time

1:28:30

it will be sorted out and eventually it will be also true for like you know generating visual arts but that does not

1:28:39

we already know that this is already not interesting since you know since the since since

1:28:46

since joshua put the fountain and you know like in the in the museum and

1:28:51

like since we had we like basically we’re like we have these people who are saying yeah we solve the

1:28:56

problem of like you know of uh of uh generating

1:29:02

good modern art i guess but like we’re not we’re all we’re not as artists we’re not uh

1:29:08

you know that it refers to a what is still a public conception of what art is

1:29:14

and like in like you know and for for most people art is like uh

1:29:19

mainstream uh well executed formal novelty that

1:29:25

changes every week and and you know i don’t see what it’s already industrialized so i don’t

1:29:32

see what we’re losing with having it you know may be be made by

1:29:38

google aside from maybe some concentration of power in the hand of the

1:29:44

companies that are already like so i i don’t i don’t think it’s a problem for arts is what i’m saying because as as

1:29:51

because it’s the fact remains that the kind of things that we’re doing uh and that a number of other people are doing

1:29:57

with these technologies is just not this approach is not to try to to it’s just to really to use these technologies

1:30:06

uh in strange ways and and and to to to open up these um these new potential so

1:30:12

i i was thinking about you were mentioning these cut ups and i i i don’t know i came to this conference thinking about ulipo you know this french

1:30:19

movement and i really like this so ulipo comes from the literature which means

1:30:25

opener of potential literature and i really like this idea of like you know these technologies being

1:30:32

openers of potentials of new ways to imagine things so

1:30:37

i was going to say that one of those contests in computational creativity one of the past the turing

1:30:43

test contest is the um there’s a generate a sonnet

1:30:48

um in order to like try to make people think like people

1:30:54

then have to pick like whether the sonnet is computer generated or human generated and i the the thing for me is

1:30:59

like okay the people who are making entries to this contest um they’re doing a lot of research

1:31:07

working with words coming up with the system for putting that we’re putting those words in order and then producing

1:31:12

those words as an artifact there’s a word for a person who does that process it’s poet like that is literally the

1:31:19

work of a poet um and from that perspective it’s like you know even though they might not think that they’re poets that’s exactly

1:31:27

what they’re doing right and the same thing goes for like composing music right like even if that process is

1:31:32

eventually automated and to a large extent it is even without like the process being like even when we’re using

1:31:39

like software to compose music we’re taking the suggestions of the software right there’s still people whose

1:31:45

responsibility it is to bring about the system that puts the notes into order on the page

1:31:51

and i don’t i don’t see like just from a purely like ontological whole

1:31:56

perspective how that person’s work differs from that of a composer right

1:32:02

i agree and i was going to think talk about computer aided creativity as opposed to computational creativity and

1:32:09

this idea that there is a system that could be a novelty generator that maybe

1:32:16

is informed by a model of your own aesthetic but it’s just there to provide suggestions or alternatives or

1:32:23

other ways of thinking about what you’re doing and not not be autonomous and then there’s always this question of

1:32:29

of sure there’s companies that are generating music but there’s always this this question of

1:32:35

at what level of description is that autonomy applied you know you could generate the notes

1:32:41

but not the music you could generate the notes and the music you could generate your own instruments that generate the music you could generate you know a song

1:32:48

you could generate an album you could generate a business model you could generate a marketing like there’s no

1:32:53

like you can’t autonomize everything because again we’re always going to get back to somebody who’s

1:32:59

orchestrating these systems yeah i would say like second per second even in like

1:33:04

a perfectly generated piece of music that would trick anybody into thinking that it’s not computer

1:33:10

generated there’s more human choice going into the into that second than like any

1:33:16

mozart sonata and maybe a lot more human labor actually a ton more human labor

1:33:23

but the promise is that uh by automating it that eventually you you’ll just be

1:33:28

able to patent the algorithm and then have just fire everyone that that and that’s a that’s a threat and i think

1:33:34

another threat and you know uh because i you know i’m kind of joking and saying that this is kind of a good news for

1:33:40

archer i don’t think it’s entirely good news for artists i think one of the bad news is that basically uh

1:33:46

so if you want to make a system that generates all kinds of genres of music and this way you don’t have to pay

1:33:52

artists anymore well the system will actually have to be trained on music

1:33:57

generated by humans you know and we know that already like well you’re all aware

1:34:02

that like artists are not like most artists are not making a lot of money they’re living in like

1:34:08

really harsh condition they work really hard and basically to have all of this production being

1:34:14

uh stolen or like appropriated by structure like structures of power that

1:34:20

are that that that that you know i think that’s that this is the there’s a there

1:34:26

is a problem there which is related to this replacement of labor or by ai but creative people are already

1:34:32

exploited right creative people are already exploited yeah but do we want them to be

1:34:37

exploited more i can’t i guess that’s true

1:34:44

[Applause] you

No results found