Tensor Decompositions for Learning Latent Variable Models

tensor-decompositionlatent-variable-modelsmethod-of-momentsspectral-methodsmachine-learning

Abstraction: AAAI 2014 tutorial on tensor methods for latent variable model learning

Key points:

Tutorial surveys method-of-moments algorithms for learning latent variable models via low-rank decompositions of higher-order tensors
Targets both practitioners using latent variable models in applications and researchers developing new learning algorithms
Prerequisites: familiarity with mixture of Gaussians and basic linear algebra and probability only
Covers techniques for developing learning algorithms based on spectral decompositions and analytical tools for understanding them
Advanced topics include learning overcomplete representations and applications to bandit problems
Presented at AAAI 2014; materials available at Columbia CS page of Daniel Hsu

Connections: Columbia University · Tensor Decomposition · Latent Variable Models · Spectral Methods · Method Of Moments