K-Deep Simplex: Deep Manifold Learning via Local Dictionaries

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Tankala, Pranay, Tasissa, Abiy, Murphy, James M., Ba, Demba
Format: Preprint
Published: 2020
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909274542702592
author Tankala, Pranay
Tasissa, Abiy
Murphy, James M.
Ba, Demba
author_facet Tankala, Pranay
Tasissa, Abiy
Murphy, James M.
Ba, Demba
contents We propose K-Deep Simplex(KDS) which, given a set of data points, learns a dictionary comprising synthetic landmarks, along with representation coefficients supported on a simplex. KDS employs a local weighted $\ell_1$ penalty that encourages each data point to represent itself as a convex combination of nearby landmarks. We solve the proposed optimization program using alternating minimization and design an efficient, interpretable autoencoder using algorithm unrolling. We theoretically analyze the proposed program by relating the weighted $\ell_1$ penalty in KDS to a weighted $\ell_0$ program. Assuming that the data are generated from a Delaunay triangulation, we prove the equivalence of the weighted $\ell_1$ and weighted $\ell_0$ programs. We further show the stability of the representation coefficients under mild geometrical assumptions. If the representation coefficients are fixed, we prove that the sub-problem of minimizing over the dictionary yields a unique solution. Further, we show that low-dimensional representations can be efficiently obtained from the covariance of the coefficient matrix. Experiments show that the algorithm is highly efficient and performs competitively on synthetic and real data sets.
format Preprint
id arxiv_https___arxiv_org_abs_2012_02134
institution arXiv
publishDate 2020
record_format arxiv
spellingShingle K-Deep Simplex: Deep Manifold Learning via Local Dictionaries
Tankala, Pranay
Tasissa, Abiy
Murphy, James M.
Ba, Demba
Machine Learning
Information Theory
Signal Processing
Optimization and Control
We propose K-Deep Simplex(KDS) which, given a set of data points, learns a dictionary comprising synthetic landmarks, along with representation coefficients supported on a simplex. KDS employs a local weighted $\ell_1$ penalty that encourages each data point to represent itself as a convex combination of nearby landmarks. We solve the proposed optimization program using alternating minimization and design an efficient, interpretable autoencoder using algorithm unrolling. We theoretically analyze the proposed program by relating the weighted $\ell_1$ penalty in KDS to a weighted $\ell_0$ program. Assuming that the data are generated from a Delaunay triangulation, we prove the equivalence of the weighted $\ell_1$ and weighted $\ell_0$ programs. We further show the stability of the representation coefficients under mild geometrical assumptions. If the representation coefficients are fixed, we prove that the sub-problem of minimizing over the dictionary yields a unique solution. Further, we show that low-dimensional representations can be efficiently obtained from the covariance of the coefficient matrix. Experiments show that the algorithm is highly efficient and performs competitively on synthetic and real data sets.
title K-Deep Simplex: Deep Manifold Learning via Local Dictionaries
topic Machine Learning
Information Theory
Signal Processing
Optimization and Control
url https://arxiv.org/abs/2012.02134