Michal S.Nov 29, 2022

AI Art from the perspective of a Machine Learning hobbyist, a thread:

There has been a lot of debate on the internet over the last year or so regarding AI art.
I think a lot of this controversy is justified, especially when it comes to datasets that ignore data protection laws.

Show threadMichal S.Nov 29, 2022
I also think a lot of the controversy around it and Machine Learning in general is FUD (Fear, Unrest, and Doubt) and comes from a wider misunderstanding of ML. Let me explain:
Show threadMichal S.Nov 29, 2022
Machine Learning is a really difficult topic to dissect, but I'll try my best from a hobbyist's perspective. To give some personal insight, I've dabbled in Machine Learning both through text-generating GANs and ones that generate images since late 2017,
Show threadMichal S.Nov 29, 2022

mostly using TensorFlow and Python.

Many people use the term “Learning” when referring to how AI is able to replicate something with increasing accuracy through ML.

Show threadMichal S.Nov 29, 2022
Many may believe this is a pretentious term and that the AI rather takes samples and picks them apart to create amalgamates of other peoples' art, and then posts that. This could not be further from the truth, and in fact most ML networks are trained to explicitly not output
Show threadMichal S.Nov 29, 2022

anything too similar to any one of its inputs.

The term “Learning” in this case, I believe, is very justified.
ML works a lot like the human brain!
You start with an input
(e.g., our eyes seeing a bear).

Show threadMichal S.Nov 29, 2022
This then is piped through any amount of hidden layers (our brain), which may have already been exposed to similar inputs and come to a conclusion
(e.g., you saw a TV show where a guy was killed by a bear, therefore, run!)
This generates an output
(In this case: “run!”)
Show threadMichal S.Nov 29, 2022
Image generation works in a very similar way.
You have a set of data
(e.g., you've seen many vases before)
Which is piped through the hidden layers
(e.g, you're an artist, so you know how to draw something you see)
In the end, this generates an output
Show threadMichal S.Nov 29, 2022

(In this case: A drawing of a vase)

Now, it's important to understand that these “Hidden Layers” are “Hidden” for a reason.
It is entirely impossible for us to understand what is going on in each node of these layers at any given point.

Show threadMichal S.Nov 29, 2022
It's entirely incomprehensible to us!
Only the AI itself understands it enough to generate an output.
This is why I make the comparison to a human learning, and why the wider field is called Machine Learning.
Show threadMichal S.Nov 29, 2022
Now, where does AI art fit into this?
We know that with AI art you can feed a prompt and have the AI generate an image, how does that work?
Show threadMichal S.Nov 29, 2022
I'm glad you asked! AI art that takes a prompt to generate works just like the image generation network I described above, however with each image, fed into it is a description (or a list of keywords).
Show threadMichal S.Nov 29, 2022
For example, a set of vases could contain “floral vase”, “ancient Egyptian vase”, “blue broken vase”, etc.
The AI then is trained to not only recreate an image of a Vase, but to understand how these keywords affect the output.
Show threadMichal S.Nov 29, 2022
The goal of this is to be able to enter a prompt that does not exist in the dataset (e.g, “floral blue vase”) and have it successfully generated.
Show threadMichal S.Nov 29, 2022
This is where AI art as we know it right now comes in. A huge problem with Machine Learning as it stands right now is there is no way to stop your art from being submitted into a Machine Learning dataset.
Show threadMichal S.
This is the root of why some AI art tools may output stolen artwork, or even in the recent case of the LAION-5B network (https://arstechnica.com/information-technology/2022/09/artist-finds-private-medical-record-photos-in-popular-ai-training-data-set/) private medical records (yikes.)
Artist finds private medical record photos in popular AI training data set

LAION scraped medical photos for AI research use. Who’s responsible for taking them down?

Ars Technica
Nov 29, 2022 at 12:32pmWeb
Show threadMichal S.Nov 29, 2022
Snore fest over, I hope that concluding this thread, you understand how AI art actually works and the real concerns over AI art.
Show threadMichal S.Nov 29, 2022
I believe one day in the future, a law should be passed that forces datasets for machine learning to be public and auditable for copyright law and data protection law infractions.
Maybe then, we'll be able to use ethical datasets for generative works, and actually help others!
Show threadMichal S.Nov 29, 2022
For example, one could use an AI to generate a very specific pose they're looking for as a reference, that's a very good use of AI and ML! I'm personally of the belief that directly using anything coming from ML/AI and calling it your own is problematic, and the laziest use of ML
Show threadMichal S.Nov 29, 2022
P.S. If you just give an AI a prompt and call yourself an Artist… I'm sorry to break it to you, bud, but you're basically just commissioning the AI…