AnchorNet is a
package and framework to perform fast data-driven low-rank approximation for dense matrices associated with
a smooth kernel function. AnchorNet takes data and kernel function as input and does not require forming the kernel matrix.
I/O illustration:
Given dataset X, kernel function f, hyperparameters ---> low-rank factors U, V for approximating the numerically low-rank kernel matrix f(X,X)
- AnchorNet is generic, allowing user-defined kernel functions and data distributions.
- AnchorNet works for datasets in arbitrary dimensions.
- AnchorNet does not require forming the kernel matrix.
- The total complexity of AnchorNet is O(rN) for computing a rank-r approximaiton to an N-by-N kernel matrix.
AnchorNet can be used whevever the kernel matrix f(X,X) is numerically low-rank, i.e. the singular values decay rapidly. If the singular values decay slowly, then it is inappropriate to use a low-rank approximation.
AnchorNet is in active development (currently 1.0.0) and its interface may change.
main.m
- Fast Deterministic Approximation of Symmetric Indefinite Kernel Matrices with High Dimensional Datasets by Difeng Cai, James Nagy, Yuanzhe Xi
AnchorNet attempts to follow semantic versioning. Do note, that in it's current (1.0.0) development, such versioning may not be strictly followed.
Difeng Cai