View in #questions-forum on Slack
@Markus: @vlomonaco is this leaderboard up to date? (https://vlomonaco.github.io/core50/leaderboard)
@vlomonaco: Unfortunately not… Maybe one day I’ll find the time, but the effort is very hard since every paper has its own setting exeptions so reporting just the acc is not significative
@Markus: Ah okay. Because I found papers where they claim to be better. Do you know a paper with one of the best results in online learning (no memory replay etc.) on Core50. Would be very interesting whats the sota accuracy value using online learning according to your settings proposed in your paper.
@vlomonaco: I’m not sure but in this paper we compare AR1 (our strategy) with DSLDA and others, achiving very good performance w/o rehearsal: https://arxiv.org/abs/1907.03799
arXiv.org: Rehearsal-Free Continual Learning over Small Non-I.I.D. Batches
@Markus: That’s what confused me, because I found your paper on AR1 (https://arxiv.org/pdf/1806.08568.pdf) and there you reach about 70% accuracy on Core50 NC. Thats much better than the leaderboard, but it does not appear there.
Thanks . I didn´t know this paper yet. Its interessting but right now I am more interessted in NC. Sorry I should have been more precisely:sweat_smile:
@vlomonaco is there a reason why AR1 is not listed on the leaderboard?
@vlomonaco: no, just I never found the time update the leaderboard!
@Markus: haha okay, thx