Split MNIST as SIT

I was reading about the characterization of CL scenarios and found that Split MNIST is considered as a Multi Task problem where each task is isolated and new classes are introduced in each task. I was wondering why it isn’t considered as a Single Incremental Task. Can’t we consider it as a single task of distinguishing between groups A and B and with subsequent batches we see new classes? An example might be that A = {0, 2, 4, 6, 8} and B = {1, 3, 5, 7, 9} and each batch contains 2 distinct classes that aren’t seen in later batches?

Hi @shrey15, you’re right that Split MNIST does not necessarily have to be a Multi Task problem. You could take a look at this paper: [1904.07734] Three scenarios for continual learning. It actually uses Split MNIST as example, illustrating that it can be performed according to different “scenarios”.

Yes, I have seen that example about Split MNIST. So, I guess Multi Task vs Single Incremental Task is implementation dependent then?

Also, for Single Incremental Task, is it mandatory for task labels to not be provided at test time? In some places I read that test time task labels are not provided for SIT (they also call it class incremental). But other places made no reference to task labels. Just wondering if test time labels are an important characterization for SIT scenarios?