Closed arthurdouillard closed 3 years ago
Hey,
Congrats on your paper, it's very interesting.
I just had a quick question, did you try longer setting such as in PODNet, with 50 steps of 1 class each for CIFAR100/ImageNet100? This is where the forgetting is the most important and thus wondered if DCE could help a lot there.
Thanks :)
I tried on CIFAR but the result was not so good as in 5 steps or 10 steps. I think the reasons are two-fold: 1) the baseline like your work is pretty well in such a setting; 2) my work tends to sacrifice novel classes and thus finds the long-run hard to handle.
Hope you enjoy your research in anti-forgetting and thank you for your help!
Thank for the info!
Hey,
Congrats on your paper, it's very interesting.
I just had a quick question, did you try longer setting such as in PODNet, with 50 steps of 1 class each for CIFAR100/ImageNet100? This is where the forgetting is the most important and thus wondered if DCE could help a lot there.
Thanks :)