Hypothesis (& one I’m pretty sure is true): adversarial examples will be a permanent problem in deep learning/ML. They exist because these systems do not have an underlying model that corresponds to their behavior.https://twitter.com/catherineols/status/1020458649825636352 …
-
Show this thread
-
The machine appears to “know” what a face is—to generalize from examples and to make human-like mistakes. Except, you can make a small perturbation that leaves the image apparently unchanged (to someone who actually knows what a face is...)
1 reply 1 retweet 3 likesShow this thread -
...and get the system to say it’s anything you want. It’s very different from an optical illusion, in the human brain, which is driven by an error in a high-level theory.
4 replies 1 retweet 5 likesShow this thread -
I believe (ought to write this up!) that it’s related to the unusual properties of sparse convex hulls in high dimensional spaces. ie, what happens when you try to find a pattern in relatively small amounts of high-d data, without a strong prior (or, as in DL, an incoherent one).
2 replies 3 retweets 17 likesShow this thread -
Replying to @SimonDeDeo
I am willing to bet a bottle of Lagavulin that ml systems will be able to deal with most adversarial examples in ten years from now. Are you willing to bet against it?
2 replies 0 retweets 0 likes -
Replying to @Plinz
Depends what you mean by most—I’m not interested in the security aspect; just the possibility of their construction. I’ll go out on a limb and say construction will always be possible except perhaps at isolated points.
1 reply 0 retweets 0 likes -
Replying to @SimonDeDeo
Lets say: I am willing to risk a bet that AI will not perform systematically worse than humans.
1 reply 0 retweets 2 likes -
Replying to @Plinz
Are we both allowed to design a stimulus for the populations in question?
1 reply 0 retweets 0 likes
At the moment, yes. No idea if the AIs will allow it in ten years from now ;)
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.