Fig. 2: Anchor points in an object manifold.
From: Separability and geometry of object manifolds in deep neural networks
Anchor points on manifolds define the optimal hyperplane separating a binary dichotomy of manifolds. This figure shows one realization of an anchor point on the ‘vase’ manifold (star) that separates it from the ‘head cabbage’ object manifold. Another realization illustrates a different anchor point on the ‘vase’ manifold when it is classified against the ‘birdhouse’ object manifold. The statistics of the distribution of these anchor points define the relevant geometrical manifold radius and manifold dimension of the ‘vase’ manifold that characterizes its linear classification properties (see Methods).
