year:
paper: intelligence-from-learnable-novelty
website:
code:
connections:
Q: Why isn’t the observer’s own description length part of the objective? Wouldn’t that let you search over observers?
Paraphrased answer by Yanbo:
Data length large enough ⇒ observer length is marginal. When observer is universal. In theory you can always translate one observers description to another. For large enough data, the lens doesn’t change enough after translating. Can always have a finite size translater between different languages!
But they are working on finding observer, cuz it does affect the result smtms.
Cifar 10 doesnt work to show clusters with random observer like with mnist. But pretrained resnet it works.
but that requires the thing to be an unbounded observer, which is the entire point of this paper is that to make it practical; and the example he then gave of cifar clustering not working with a random redout but yes working with a pretrained resnet… shows this even more? anyways.
Q: With a frozen observer, isn’t this a fixed objective… where does the open-endedness come from?
Some interesting brainstorming: e344ced1-bf81-4700-9a9d-f160ee8d2754