Is Subjective Beauty Something We Can Model with AI?

How does the human brain decide someone is beautiful? It is something universal or subjective? This AI attempts to help give us some answers to these questions.

This AI reads your brain to generate personally attractive faces. It generates images containing optimal values for personal attractive features on a human face!

Watch the video

References

DeepFakes in 5 minutes:
GANs explained:
An amazing application of GANs:
Read the article: https://whats-ai.medium.com/

M. Spape, K. Davis, L. Kangassalo, N. Ravaja, Z. Sovijarvi-Spape and T.
Ruotsalo, "Brain-computer interface for generating personally attractive
images," in IEEE Transactions on Affective Computing, doi:
10.1109/TAFFC.2021.3059043.

Abstract—

While we instantaneously recognize a face as attractive, it is
much harder to explain what exactly defines personal attraction. This
suggests that attraction depends on implicit processing of complex,
culturally and individually defined features.

Generative adversarial neural networks (GANs), which learn to mimic complex data distributions, can potentially model subjective preferences unconstrained by pre-defined model parameterization. Here, we present generative brain-computer interfaces (GBCI), coupling GANs with brain-computer interfaces.

GBCI first presents a selection of images and captures personalized attractiveness reactions toward the images via electroencephalography. These reactions are then used to control a GAN model, finding a representation that matches the features constituting an attractive image for an individual.

We conducted an experiment (N=30) to validate GBCI using a face-generating GAN and producing images that are hypothesized to be individually attractive. In double-blind evaluation of the GBCI-produced images against matched controls, we found GBCI yielded highly accurate results.

Thus, the use of EEG responses to control a GAN presents a valid tool for interactive information-generation. Furthermore, the GBCI-derived images visually replicated known effects from social neuroscience, suggesting that the individually responsive, generative nature of GBCI provides a powerful, new tool in mapping individual differences and visualizing cognitive-affective processing.

Video Transcript:

00:00

i think you will all agree that when you

00:02

look at someone

00:03

and find the person attractive you

00:05

cannot explain why

00:07

there are many reasons involved in this

00:09

decision

00:10

plus some of these reasons may not even

00:12

be directly related to the physique of

00:14

the person

00:15

still when you look at a picture of

00:17

someone you instantly know if the person

00:19

is attractive or not by your standards

00:21

and you know this without any doubts but

00:24

how does that happen

00:26

can we explain or understand what beauty

00:28

is to us and worse can someone else

00:31

understand what we find beautiful is

00:34

beauty even something universal

00:36

or subjective to me it's definitely

00:39

something subjective

00:40

just like art as we say beauty is in the

00:43

eye of the beholder

00:44

isn't it but then why specific people

00:47

become

00:48

actors or models and not others is there

00:51

a universal beauty that we could explain

00:53

is beauty something cultural biological

00:56

or a blend of both

00:58

i think these are mostly philosophical

01:00

debates that we can discuss in the

01:01

comments

01:02

for those of you that will like to i

01:04

would love to chat and even debate about

01:06

this with you

01:07

anyways back to what we know about

01:09

beauty as we know

01:10

humans typically respond emotionally to

01:13

attractive images

01:14

rather than purely based on rational

01:16

visually salient reasons

01:18

which is what makes beauty so subjective

01:20

but is there a way to understand what we

01:22

think

01:23

of beauty and use this information to

01:25

create attractive pictures

01:27

can we understand our brain signals to

01:29

achieve that

01:30

and is it even readable to these signals

01:33

well

01:34

this is what michel spape and his team

01:36

from the university of helsinki

01:38

attempted to do what you see here is

01:41

called electroencephalography

01:43

or eeg it's a monitoring method used to

01:46

record electrical brain activity

01:48

using electrodes placed along the scalp

01:51

specifically they tracked

01:52

two different channels with these

01:54

electrodes which are known to be

01:56

responsive to stimuli that are either

01:58

novel or effectively evocative or

02:01

mentally demanding and related to

02:03

working memory updating

02:05

meaning that the stimuli are improbable

02:07

and require a mental or physical

02:09

response

02:10

but they employed two strategies in

02:11

their experiments

02:13

first they made the supposed attractive

02:15

images relatively improbable to appear

02:18

they also asked the participants to

02:20

focus on attractive faces

02:21

by mentally counting their occurrence to

02:24

me these strategies seem like a

02:26

considerable bias to detecting

02:28

actual attractiveness where the signal

02:30

responses mainly tell us two things

02:33

one that the person is lightly surprised

02:35

because the attractive pictures are rare

02:37

and two that the person is mentally

02:39

active trying to remember the count and

02:41

add to it

02:42

i personally see it as an overkill way

02:45

of asking whether the person finds the

02:46

face attractive or not

02:48

since it does not really measure the

02:50

beauty of the image scene

02:52

and here i am not entering about the

02:54

details of the data set they used to

02:56

train their model

02:57

which is only made of celebrities

02:59

nonetheless

03:00

this was undoubtedly a great way to find

03:03

out

03:03

which faces were attractive or not to

03:06

train their model to focus on beautiful

03:08

faces

03:08

which is a very cool application of guns

03:11

as you can see it clearly worked out

03:13

there's a significant difference between

03:15

responses after a slight delay of time

03:17

starting approximately a fourth of a

03:19

second after showing the image

03:21

where the pink lines are the two channel

03:24

responses i just mentioned

03:25

are shown here the top shows the frontal

03:28

response

03:29

responsible for the novel and

03:31

effectively evocative stimuli while the

03:33

bottom is the parietal response

03:35

responsible for mentally demanding

03:37

stimuli here

03:38

the faces deemed attractive are green

03:40

unattractive in red

03:42

and inconsistent responses in gray these

03:45

inconsistent responses signified that

03:47

the data

03:48

was not used due to its low level of

03:50

confidence

03:51

attractiveness was confident when both

03:53

electrodes evoked more positivity for

03:55

the attractive faces

03:57

now let's see how they created such a

03:59

model to use their brain activity to

04:01

generate attractive faces

04:03

at first they needed to generate random

04:05

images

04:06

to show the test subjects to do that

04:08

they used a gun model trained on 200 000

04:11

images of celebrity faces already

04:14

introducing a bias

04:15

but still let's keep going they then had

04:18

a latent space

04:19

from which they could use to generate

04:21

new artificial images

04:22

this is just like any other gun model

04:25

like deep fakes

04:26

it is very similar to what i already

04:28

explained in previous videos

04:29

and you are most certainly aware of how

04:31

it works so i won't cover it again here

04:34

but if you are not

04:35

you are free to click on the top right

04:37

pop-up and watch my explanations of

04:39

gan architectures and i'll talk a little

04:41

bit more about this latent space

04:43

later on then each participant was asked

04:47

to go through

04:47

randomly generated images remember here

04:51

they had to count when they see an

04:52

attractive face mentally

04:54

once the results are all compiled they

04:56

train a classifier on these brain output

04:59

signals

04:59

containing the image's subject

05:01

assessments

05:02

basically ideally containing whether the

05:05

face was attractive or not

05:07

once the classifier is trained it's used

05:09

under brain electrical responses

05:11

on randomly generated images note that

05:14

all these randomly generated images come

05:17

from a latent space originally

05:19

created by the encoder of the gun

05:21

trained on the first step

05:22

as you can see here the latent space is

05:25

a space where one point represents

05:27

an input the generator can use to create

05:29

a unique image

05:30

this space contains all the faces that

05:33

this generator can make

05:35

thus when an image is generated using a

05:37

point in the space

05:38

or feature vector it can be considered

05:41

attractive or not

05:42

by the classifier if that's the case it

05:44

will encourage the network to create

05:46

images by selecting the generator's

05:48

input close to this point in the space

05:50

focusing the generator's attention on a

05:52

subspace

05:53

where the features of the generated

05:55

faces are pleasing to the participant

05:58

finally this is how new images

06:00

containing optimal values for personal

06:02

attractive features on a human face

06:05

are generated i hope you enjoyed this

06:07

video

06:08

and i definitely invite you to read

06:09

their paper it's very interesting and

06:12

well written it's the first link in the

06:14

description don't forget to like and

06:15

subscribe if you enjoyed the video

06:17

and share your thoughts in the comments

06:19

where i will answer you in the following

06:21

7 minutes

06:22

i promise thank you for watching