Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
INRIA LMO, University of Warwick, Université Paris-Saclay
1
0
2
ReLU networks trained with gradient flow on orthogonal data exhibit an incremental learning process, activating neurons one-by-one at saddle points, and converging to interpolators with near-minimal $\ell_2$ norm.