Connectionists: New code release: Locally Linear Landmarks (LLL) for manifold learning

[Max Vladymyrov]

Dear all,

We are pleased to advertise Matlab code for computing Locally Linear 
Landmarks (LLL) as described in the following paper:

M. Vladymyrov and M. A. Carreira-Perpinan: "Locally Linear Landmarks for 
Large-Scale Manifold Learning", ECML-PKDD 2013, pp. 256-271.
https://eng.ucmerced.edu/people/vladymyrov
http://faculty.ucmerced.edu/mcarreira-perpinan/papers.html

The code computes a fast, approximate solution to Laplacian Eigenmaps and 
potentially any other spectral method. 

Laplacian Eigenmaps is a very popular nonlinear dimensionality reduction
method. However, it requires computing the trailing eigenvectors of an NxN
positive semidefinite matrix, which is expensive with large N. LLL solves
a reduced eigenproblem by constructing a new affinity matrix for a subset
of landmark points so that the manifold structure is better preserved than
with Nyström's method. For datasets with millions of points, this results
in significant speedups with only a small approximation error.

Sincerely,
Max Vladymyrov and Miguel A. Carreira-Perpinan