Questions about loss

[Kiljoon]

Thanks for the great work.

The "fixed-point correction" appears to be applied in a dense manner, as seen in RAFT and similar methods. However, the paper mentions that it is applied in a sparse manner. What are the differences in the results between these two approaches?

Furthermore, it seems that intermediate hidden states need to be stored to compute the fixed-point correction. However, changing the f_thres did not result in any difference in memory usage. Can you explain why this is the case?