View in #questions-forum on Slack
@Kostis_SZ: Hey everyone! I ve been reading the paper “Continual learning: a comparative study on how to defy forgetting in classification tasks” arxiv.org/pdf/1909.08383.pdf and I found interesting (although maybe a bit too abstract?) their summary of the 10 desiderata of an ideal continual learning scheme should follow. Do you have any thoughts / agreements / disagreements with those? (I am still quite new to this field so I am trying to understand what are some ideas the community agrees / disagrees on)
@vlomonaco: @Rahaf maybe you want to answer this? 
Personally I agree with all of those. Actually I’ve been pushing for years in this direction (see my works): I’m very happy to see many other researchers are now working with this desiderata in mind and not being limited by the standard Multi-Task setting!
@Rahaf: Yes these desiderata are gathering from discussions with other researchers in the field and reading papers like yours. Personally I think those are the characteristics to take into account and thus I had them in my thesis ;). It would be nice to see what other people think
@Kostis_SZ: Thank you for your answer, I wasn’t aware this slack contained authors of this paper! 
Could I ask for a bit of clarification between the desideratum 5 (backward transfer) and desideratum 9 (task revisiting)? Would you agree if I rephrased 9 as “new task knowledge should be leveraged when evaluating on previous tasks”? But then doesn’t this sound a bit similar with backward transfer? Is there a distinction I am missing here?