Deep Random Feature Architopes
This code create a partition of the input space by:
- Generating a large number of homeomorphic feed-forward networks,
- Determine which top N feature maps improve performance most,
- Train a classifier
$s:\mathbb{R}^d \rightarrow {1,\dots,N}$ to predict which random feature map works best, - Define
$K_n\triangleq s^{-1}(n)\cap [-M,M]^d$ , for some large$M>0$ .