Question about applicability

[UpasanaSridhar]

Hi, I came across your paper a few weeks ago.
I have a dataset that constantly grows, like every couple of weeks. The growth is both in terms of more examples of a set of known classes as well as new classes being added.
Is it possible to use this method to keep a reduced dataset of the old images?
For eg: I have 10k images that I want to distill into 100. Then I get a new batch of 200 images.
How would I retrain a model "from scratch" using this combination of distilled and raw images?

I'm a grad student focussing on HPC, so I'm sorry if these questions are silly. But I would greatly appreciate any feedback, thank you!