Search papers, labs, and topics across Lattice.
3
0
3
SVRG, a popular variance reduction method, finally gets its generalization guarantees rigorously established, revealing optimal excess population risk bounds.
Momentum doesn't hurt generalization in SGDM, and this paper proves it with tight stability bounds applicable to both Polyak's and Nesterov's variants.
Overparameterized neural nets can generalize well even with many parameters, and this work provides tighter, initialization-aware bounds that scale logarithmically with network width, offering a more nuanced understanding of this phenomenon.
