@ngutten

Physicist/ML researcher creating artificial universes. @nichg@sigmoid.social

Joined November 2012
  • Tweets7.4K
  • Following241
  • Followers2.1K
861Media
Experiments with #stablediffusion... Its very hard to get it to generate images with average value outside of the 0.4..0.6 range, even if you fine-tune on a solid black image. Turns out its a property of the noise distribution. i.i.d 512x512 Gaussian has very close to zero mean.
1
7
811
Nicholas Guttenberg retweeted
Louis Kirsch@LouisKirschAI2 Dec 2022
How can Transformers and other nets implement general-purpose learning algorithms in-context? Training on 2^14 tasks leads to strong generalization. Diff. to scaling laws, the memory size matters more than parameter count! Meta-Learn Workshop #NeurIPS2022 louiskirsch.com/gpicl
4
47
3
197
Nicholas Guttenberg retweeted
Kevin Frans
@kvfrans26 Nov 2022
Powderworld is a lightweight simulation environment for understanding AI generalization, supporting: - modular & emergent ruleset - expressive RL task space - 10k timesteps/sec via GPU Paper: arxiv.org/abs/2211.13051 Blog + Online Demo: kvfrans.com/static/powder/
13
115
24
624
Interesting dynamics of inference... Imagine a coin flip process, except you entertain hypotheses that after 1,2,3,...,K flips the sequence will repeat. This is the distribution of predictions that the next flip will be tails as you observe more and more for K=1000
1
1
8
more replies
So the cluster would have the biggest impact on expected values of things just before you stop believing in it entirely...
1
1
In particular, I guess you'd need past observations up to this point to have a less correlated effect on 'adjacent' hypotheses than those hypotheses' impacts on future predictions. So you shouldn't see it in stationary processes I guess?
Is there such a thing as an 'addictive experiment' in science: an experiment which no matter how many times it has been repeated, never seems unimportant to repeat again? (e.g. remains asymptotically interesting in the limit of infinite observations)
4
1
6
more replies
Hm, maybe this works. Imagine a sequence of heads or tails. Base model space might be an iid biased coin with beliefs over the bias of the coin. Now add to that model space 'repeating sequences of length N where elements of the sequence are biased flips' for all values of N
3
I don't know that it quite works with the KLdiv here because its an infinite hypothesis space, so all of the probabilities of individual hypotheses are infinitessimals...
Messing with the LMS scheduler in Stable Diffusion. One trick I've found is to use scheduler.add_noise to do partial re-diffusions, e.g. starting halfway through given a starter image. So I can basically let it drift an image by iteratively rediffusing.
1
9
more replies
Another one, lower equilibrium standard deviation of the latent.
1
4
Here I'm using a mask to determine where the noise vector is constant (edges) and where it moves (center). Maintaining proper normalization is pretty tricky here, I had to fiddle with it by hand to find a multiplier that kept things relatively stable.
5
Any reason why Stable Diffusion style autoencoder + diffusion on latents couldn't be used for very long self-consistent text generation - like 30k tokens+? Have people already done this?
9
With content that is less naturally symmetric, the transition is even more severe...
1
When SD decides 'these repeated noise patches indicate real structure', its a really sharp phase transition (as one might expect given that its a discrete symmetry). Wonder if there are 2nd order phase transitions in the diffusion model as well?
3
Load more