I love these examples too, but there's something funny to me about describing these results as "fooling" DL (as authors suggest). These images are...extremely weird. Why does DL need to call an impossibly diagonal, cropped, photoshopped vehicle a "firetruck" to impress us?https://twitter.com/GaryMarcus/status/1068169331832315906 …
-
Show this thread
-
Replying to @chazfirestone
Either I have misunderstood your tweet or you have not read the paper carefully
1 reply 0 retweets 0 likes -
Replying to @GaryMarcus
The former, I think. But maybe because I misunderstood your initial tweet? We should definitely be alarmed by DL's performance on column (d). Should we also be alarmed by DL's performance on columns (b) and (c)? The paper itself focuses much more on those, hence my tweet.
2 replies 0 retweets 0 likes -
Replying to @chazfirestone
Column D is what matters; B and C were steps to getting there (as explained in the paper)
1 reply 0 retweets 0 likes -
Replying to @GaryMarcus
Huh, that's really not what I'm getting from it. They clearly (to me at least) emphasize *both* the standalone importance of the renders and also the ability to find natural images that behave similarly. I may be reading wrong, but this summary is not about natural images, is it?pic.twitter.com/IE7ZldOWuR
1 reply 0 retweets 0 likes -
Replying to @chazfirestone
I will post my interpretation tomorrow morning on Medium; meanwhile, looping in
@anh_ng81 reply 0 retweets 0 likes -
Replying to @GaryMarcus @anh_ng8
Great. (Hey
@anh_ng8! Hope you're well)2 replies 0 retweets 1 like
Started new thread with tweet linkinghttps://link.medium.com/2FUz9lTsiS
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.