François CholletVerified account

Smells a bit fishy to me too. arXiv submission only after NIPS acceptance feels like an end run around reproduction efforts.

I'd love to see a reusable OS implementation. I want to believe

shouldn't this be very easy to reimplement though?

Indeed, and I will gladly eat my words if it turns out to be completely reproducible.

But if I had a bombshell like this, I wouldn't have sit on it for 8 weeks of NIPS reviewing; makes me suspicious.

The result isn't that surprising is it? Roughly stacked sparse coding. Similar models have been built before

Only difference is claim to match VGGnet. But sparse-coding has given strong numbers in past

Layer-local greedy training doing that well is pretty surprising, yes.