Loss function used to train baseline classifier g about m2m HOT 1 CLOSED

Hi, thanks for interest in our work!

As we have noted in the paper, we used a vanilla classifier, i.e., trained with vanilla CE loss without mitigating imbalance as g. The training script is already given in our "run.sh" file.

In our experiments, more sophisticated losses do not improve the performance of our algorithm, M2m.
Because capturing the features of the majority class is be more important for the baseline classifier g as one can see in Table 3.
Here, it is verified that the diversity of seed samples (majority class) directly affects the improvement from M2m.

But, if vanilla CE can't learn the training dataset with high accuracy, then I recommend 1) to increase the capacity of the network or 2) try sophisticated losses. Although the later one was not effective in our case, it can be different in your case.

Best,
Jaehyung.

from m2m.