Fast Margin Maximization via Dual Acceleration
Ziwei Ji 1 Nathan Srebro 2 Matus Telgarsky 1
Abstract 0.05
We present and analyze a momentum-based gra- 0.00
dient method for training linear classifiers with 0.05
an exponentially-tailed loss (e.g., the exponential
0.10
or logistic loss), which maximizes the classifica-
2
tion margi ...
附件列表