search

VICO-UoE / DatasetCondensation

Dataset Condensation (ICLR21 and ICML21)
MIT License
487 stars 93 forks source link

Question about network weights update in Algorithm 1 #19

Closed minhquoc0712 closed 1 year ago

minhquoc0712 commented 1 year ago

Thank you for your amazing work.

Can I ask about line 9 on your Algorithm 1 in your paper? Why do you choose to update network weights based on loss on synthetic data instead of real data?

PatrickZH commented 1 year ago

Hi, thanks. Please refer to the Line 7 in the first paragraph of Page 5 in our paper https://openreview.net/pdf?id=94YYh974g3T Why? Because:

  1. empirically, it works better.
  2. the objective of dataset condensation is training models on synthetic data, thus updating models on synthetic data is reasonable and in line with the objective.
  3. recently some works such as IDC modify it by updating models on real data and achieve good results.
minhquoc0712 commented 1 year ago

I think I have gained more understanding now. Thank you for your respond.