Fast Variational Block-Sparse Bayesian Learning

1 Jun 2023  ·  Jakob Möderl, Franz Pernkopf, Klaus Witrisal, Erik Leitinger ·

We present a fast update rule for variational block-sparse Bayesian learning (SBL) methods. Based on a variational Bayesian approximation, we show that iterative updates of probability density functions (PDFs) of the prior precisions and weights can be expressed as a nonlinear first-order recurrence from one estimate of the parameters of the proxy PDFs to the next. In particular, for commonly used prior PDFs such as Jeffrey's prior, the recurrence relation turns out to be a strictly increasing rational function. This property is the basis for two important analytical results. First, the determination of fixed points by solving for the roots of a polynomial. Second, the determination of the limit of the prior precision after an infinite sequence of update steps. These results are combined into a simplified single-step check for convergence/divergence of each prior precision. Consequently, our proposed criterion significantly reduces the computational complexity of the variational block-SBL algorithm, leading to a remarkable two orders of magnitude improvement in convergence speed shown by simulations. Moreover, the criterion provides valuable insights into the sparsity of the estimators obtained by different prior choices.

PDF Abstract

Datasets


  Add Datasets introduced or used in this paper

Results from the Paper


  Submit results from this paper to get state-of-the-art GitHub badges and help the community compare results to other papers.

Methods