Is the MOON algorithm wrong

[JY132]

First of all, I need to declare that I am newbie in this field. Thus, my understanding may be wrong. If you found any issue, please let me know.

I look up to the codes posted by the authors that proposed the MOON algorithm.

The codes allow multiple previous models for contrastive loss as shown in the figure below.

However, in your implementation, the local training only involved single old client' model.

[TsingZ0]

In the paper, the author exclusively utilizes a single old client's model. During a conversation with the author, he mentioned experimenting with multiple old client's models and concluded that employing the last old client's model yielded the best results.