Learning latent variable models by improving spectral solutions with exterior point method

摘要: Probabilistic latent-variable models are a fundamental tool in statistics and machine learning. Despite their widespread use, identifying the parameters of basic latent variable continues to be an extremely challenging problem. Traditional maximum likelihood-based learning algorithms find valid parameters, but suffer from high computational cost, slow convergence, local optima. In contrast, recently developed spectral computationally efficient provide strong statistical guarantees, not guaranteed parameters. this work, we introduce two-stage algorithm for models. We first use method moments solution that is close optimal necessarily set model then incrementally refine via exterior point until optima arbitrarily near found. perform several experiments on synthetic real-world data show our approach more accurate than previous especially when training limited.