dATE
10 August 2025
Credit
MIT Computer Science & Artificial Intelligence Laboratory (CSAIL)
#
Generative AI, noise patterns, training data, datasets
Learning to See by Looking at Noise
https://arxiv.org/pdf/2106.05963
This paper is from before the big generative AI boom.
It trains models by feeding it randomly generated noise images as the training data itself. The noise is fixed per training example (so the same “noise image” corresponds to the same label during training). The models are not learning a denoising process.
The noise element here only concerns the input! Which I interpret as doing away with the mimetic dynamic… which leaves us with just the technological function, that is normally obscured to us because we're focused on how passable the image result is. We rarely attend to how the model is transforming information.
dATE
10 August 2025
Credit
MIT Computer Science & Artificial Intelligence Laboratory (CSAIL)
#
Generative AI, noise patterns, training data, datasets
Learning to See by Looking at Noise
https://arxiv.org/pdf/2106.05963
This paper is from before the big generative AI boom.
It trains models by feeding it randomly generated noise images as the training data itself. The noise is fixed per training example (so the same “noise image” corresponds to the same label during training). The models are not learning a denoising process.
The noise element here only concerns the input! Which I interpret as doing away with the mimetic dynamic… which leaves us with just the technological function, that is normally obscured to us because we're focused on how passable the image result is. We rarely attend to how the model is transforming information.
