Question about expected finetuning performance

[VGrondin]

Hello, I read the paper and the performances reported using linear probing are impressive.
My question is: if I were to finetune the model on iNat18 using a linear head, in your experience would the performances be better than a model pretrained on MAE, MAWS, etc?

Thanks and congrats for the work!