lsh0520/RGCL

Details of the training of the rationale generator

Closed this issue · 2 comments

Thank you for your enlightening work. I am curious about the training of the rationale generator which does not seem to be detailedly mentioned in the paper. In section 4.1, it seems that the rationale generator is trained unsupervised, but popular IRD training methods need labels. Could you tell me how to train the rationale generator through unsupervised training? Thank you.

Thank for your interest in our work.

Indeed, typical IRD trains in a supvervised fashion, while we borrowed its idea and tried to adapt it to the contrastive learning -- unsupervised learning.

Ovearlly, the gradient signal is generated by the the loss function as depicted in Equations(16-18) to perform instance-discrimination task (unsupervised learning). During this process, both the backbone encoder and the rationale generator is optimized to learn invariant features.

And the key point here is how to enable the gradient back propagation to the rationale generator. Please see Equation (13) for details.

Thank you. I have got the main idea of the rationale generator's training.