Search papers, labs, and topics across Lattice.
1
0
1
SAM's implicit bias in deep linear networks flips the script on feature learning, prioritizing minor data coordinates early in training before amplifying major ones, a behavior unseen in gradient descent.
