Some questions about CHIEF pre-training

[KKIverson]

Congratulations on this work on CHIEF.
I have a few questions about pre-training. I noticed that the supplementary file mentioned that 3 losses were used when pre-training the CHIEF model. Formula (3) is used to calculate the bag-level classification loss, and formula (4) is used to calculate the instance-level top-k loss. I understand both 3 and 4. But I don’t understand the loss in formula (5).
First, there are two negative signs in formula 5. Is there a typo (should there be only 1 negative sign) ? Second, for each bag(G), whether it is negative or positive, does the numerator calculate the similarity between each instance in G and the instance in the positive bag? If bag(G) is a negative sample, will it be counterproductive? Because your goal is to maximize the intra-class similarity and minimize the inter-class similarity.
Thank you!

[KKIverson]

Another small question: How big is this memory bank? Does it contain all the samples in this tissue source site? How is it maintained? Is it updated in each iteration?