Inference and Learning in a Latent Variable Model for Beta Distributed Interval Data.

摘要: Latent Variable Models (LVMs) are well established tools to accomplish a range of different data processing tasks. Applications exploit the ability LVMs identify latent structure in order improve (e.g., through denoising) or estimate relation between causes and measurements medical data. In latter case, form noisy-OR Bayes nets represent standard approach relate binary latents (which diseases) observables symptoms). with representation for symptoms may be perceived as coarse approximation, however. practice, real disease can from absent over mild intermediate very severe. Therefore, using diseases/symptoms relations motivation, we here ask how generalized incorporate continuous observables, e.g., variables that model symptom severity an interval healthy pathological. This transition poses number challenges including Bernoulli Beta distribution statistics. While noisy-OR-like approaches constrained determine observables' mean values, use distributions additionally provides (and also requires) variances. To meet emerging when generalizing distributed investigate novel LVM uses maximum non-linearity means variances observables. Given goal likelihood maximization, then leverage recent theoretical results derive Expectation Maximization (EM) algorithm suggested LVM. We further show variational EM used efficiently scale large networks. Experimental finally illustrate efficacy proposed both synthetic sets. Importantly, produces reliable estimating proofs concepts first tests based on images.