I’m a bit confused about how the training is performed when the coreset is initially empty

[kimjisoo12]

Thank you for your excellent work. I really appreciate the clarity and depth of the methodology presented.

I have a question regarding Algorithm 1 in the paper. In the first step, the coreset is initialized as an empty set. Could you please clarify how the model is trained using the "current subset" at this stage, given that the coreset is empty at the beginning?

Looking forward to your clarification, and thank you again for the insightful contribution.

[RuilinTong]

Thank you for your thoughtful question. The coreset is initialized with an empty set, and the initial coreset model is randomly initialized, meaning that nothing is learned by coreset model.

Assuming all samples have similar loss on initial coreset model, the first step selects samples with lower holdout model loss.