View in #questions-forum on Slack
@Martin_Mundt: Hey all, does anyone know if there are any papers investigating the role of task ordering in class incremental/continual learning and could give me pointers?
@Julio_Hurtado: I haven’t found, but it’s something that interests me. There is this paper https://arxiv.org/pdf/2001.08714.pdf that makes experiments putting the bigger task at the beginning, but it is not the focus of the work. Are you tihinking in sorting the task by complexity or some other criteria?
@Martin_Mundt: Yes, I have done some experiments with some internal complexity criteria and now I am in the process of writing these up. Naturally, part of that would be to compare to anything that already exists, but I am having a hard time finding much
@vlomonaco: Hi @Martin_Mundt! This ia good point. In our papers we always report standard deviation computed over 10 runs where we shuffle the batch order!
But it is true that the batch complexity is not that different. @Qi_(Roger)_She proposed the OpenLoris Dataset also to study this problem introducing different levels of complexity for each element of variation!
P.s. maybe you can look into the curriculum learning field and take some inspiration from there! it would be fun to finally merge it with Continual Learning. I know @German_I._Parisi and @Chris_Kanan are fan of this https://arxiv.org/abs/1907.01929
arXiv.org: OpenLORIS-Object: A Robotic Vision Dataset and Benchmark for…
arXiv.org: Rethinking Continual Learning for Autonomous Agents and Robots
@Chris_Kanan: @Martin_Mundt We have looked a little bit at order, but I don’t know if it gets at what you are interested in. See Table 1 of this paper: https://arxiv.org/abs/1910.02509
arXiv.org: REMIND Your Neural Network to Prevent Catastrophic Forgetting