Search papers, labs, and topics across Lattice.
Google
1
0
3
Randomly masking parameter updates in RMSProp delivers state-of-the-art LLM training performance, revealing a surprisingly effective form of geometric regularization.