Skip to content

Questions about Reverse Self-Contrastive Loss and LoRA Training #5

Description

@kojunseok0217

Hello, thank you for sharing this interesting work.
I have two questions that I could not fully understand from the paper:

  1. Reverse Self-Contrastive (RSC) Loss

I don’t quite understand why the RSC loss should behave differently from the conventional self-contrastive loss.

In Equation (8), if we move the minus sign inside the log, it seems to become identical to Equation (9). Doesn’t this mean they are essentially the same loss function?

Image Image
  1. LoRA Training Procedure

I would like to confirm if my understanding is correct:

  • Is LoRA trained every time a new prompt 𝑐 is given?

  • When applying the same concept to different prompt 𝑐, do we need to retrain LoRA for each 𝑐?

  • If we want to remove multiple concepts for the same 𝑐, should we (a) train multiple LoRAs at once, or (b) train a separate LoRA for each concept and then apply all of them together at inference time?

I am not fully confident about how LoRA is used in this context, so I would appreciate a clear explanation.

Thank you very much for your help!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions