GP-Select: Accelerating EM Using Adaptive Subspace Preselection.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Statistical Modeling

Background:

• Generative graphical models are powerful tools for representing complex data distributions.
• Inference in models with a large number of latent states is computationally challenging.
• Exact inference methods become infeasible when the number of latent states grows exponentially.

Purpose of the Study:

• To propose a nonparametric procedure for fast inference in generative graphical models with numerous latent states.
• To develop an automated method for selecting relevant latent variables, avoiding manual design.
• To enable computationally feasible inference for previously intractable models.

Main Methods:

• Iterative latent variable preselection: alternating between learning a selection function and approximating the posterior distribution.
• Gaussian process regression is used to learn the selection function from observed data and current expectation-maximization state.
• The learned selection function guides the approximation of the posterior distribution for efficient expectation-maximization.

Main Results:

• The proposed method achieves fast inference comparable to bespoke, manually designed selection functions.
• Demonstrated effectiveness across a variety of inference problems.
• Achieved state-of-the-art results in hierarchical object localization with occlusion at significantly lower computational cost.

Conclusions:

• The nonparametric iterative latent variable preselection offers an efficient and automated approach to inference in complex graphical models.
• This method significantly reduces computational burden, making advanced models more accessible.
• The approach generalizes well and matches specialized methods in challenging real-world applications.