CL problems / desiderata

Hi everyone!
I would to ask for more information about the others CL problems/desiderata that we want to address (in particular slide 20 in Lec 2 not only CF).

  • Backward and Forward transfer
  • Compositionality
  • Robustness


1 Like

I have a second question different from the previous one.
From permuted_and_split_mnist.ipynb lab, on the question
“* Can curriculum learning improve our final performance?”
In my opinion and in short yes

And for this reason, to better understand CL, what are the main differences with curriculum learning? (question after a quick read of the paper “On The Power of Curriculum Learning in Training Deep Networks”)

Hi @semola! I think this will be much more clear after the 4th lecture which is gonna be all about evaluation protocols and metrics!

In Curriculum Learning you mostly assume to have full control over the generation of the stream of data that you let the algorithm to process. So, based on some domain knowledge you can create a “curriculum” of increasingly more complex knowledge in order to make the algorithm learn faster and with better generalization capabilities.
In a Continual Learning setting you don’t have any control over the stream of data and/or its data sources. Hence, you cannot craft any curriculum if not with some advanced (and possibly inefficient) replay strategy.

1 Like