Playing Around with Noise as Targets

I didn't read the referenced 2017 paper yet, but mapping the training data to noise (gaussian and/or other) is exactly what the RevNet paper does, with the advantage of deterministic reversibility such that the trained RevNet is also generative (without having to do gradient descent for each generated image)

> looks like something genuinely new

I fear those words

I think I'm missing the point?

A visual proof that neural nets can compute any function: http://neuralnetworksanddeeplearning.com/chap4.html

Don't trust any machine learning algorithm that you haven't faked yourself. You can make random noise mean anything you want.

His stated purpose was image compression (although I didn’t see evidence that it worked). If the distribution encoded by the is smaller than your image, then you can send the small set of model parameters (instead of the image) and then use the model to reconstruct the target image.