daniel-gallo / naraim

3 stars 0 forks source link

Check whether performance of the probe keeps improving even if the pre-training is overfitting #18

Open robertdvdk opened 5 months ago

robertdvdk commented 5 months ago

We observe that the pre-training is overfitting, which is why we propose to add dropout. However, it might be the case that the network's representations are still improving. We should check whether the fine-tuning performance keeps improving even if the pre-training is overfitting.