Margin evolution during training

methods / Theoretical / Analytical / Geometric analysis / Margin evolution during training

Techniqueintermediate

Tracks a classifier's DeepFool-estimated margin (l2/l-infinity distance to the decision boundary) at successive training checkpoints, rather than measuring the boundary's shape only at a single fixed (fully-trained) state -- revealing that the boundary continues moving closer to training/test points well into the low-error regime, and that adversarial training reverses this trend.

Used in (1 observation)

Decision-boundary margins keep shrinking past the error plateaumeasured

structure: Decision boundary (as a codimension-1 hypersurface) · models: Dense/fully-connected classifier (MNIST), Convolutional classifier (Fashion-MNIST) · paper: Understanding the Decision Boundary of Deep Neural Networks: An Empirical Study