question about the holdout model

[LightZH]

Hi, I have a question about the holdout model.
Why retrain a holdout model instead of directly using the trained classification model?

[RuilinTong]

Thank you for your question.

In offline continual learning, we train holdout model after each task. This is based on the considerations that: continual learning model may not be optimal, and this will affect coreset selection, and coreset selection may affect later tasks. To mitigate this problem, we choose to train an independent holdout model.

For CSReL-RS, we use continual learning model as holdout model.

[LLZQYF]

For CSReL-RS, the continual learning model is used as holdout model, which is the coreset model and how does it train?
thanks

[RuilinTong]

Thank you for your question.

In CSReL-RS, the coreset model is trained from buffer data of current task, to encode what knowledge is contained in current buffer, and used for computing reducible loss. This model is trained with few steps using the same backbone model.