Comment by yorwba

6 hours ago

What is classic about "skip updating parameters with high gradient/loss variance in multiple batches/samples"? Do you have a particular algorithm in mind that uses this heuristic?

0 comments