a fair comparison with randomly shuffle all the 20 task data?

[heimanba89]

Hi,

Nice try!
I have a concern about your implementation. Basically, your current implementation have to (partially) store previous task samples in the memory, which could be problematic if the class number is big (for example 10k class).

Do you have a fair comparison with an alternative training process? I.e., shuffling the data of all the tasks, and training the network, and test on all 20 tasks? If we can see the benefits from RMA still, then we can say RMA is effective.

Chunlei