Nonlinear spike-and-slab sparse coding for interpretable image encoding

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Sparse coding models natural images using linear superposition of dictionary elements.
• Traditional models face challenges with low-level image structures (e.g., occlusions) and varying pixel intensities.
• Existing probabilistic models often use Laplace or Cauchy priors, limiting representation of exact zeros.

Purpose of the Study:

• To propose a novel sparse coding model addressing limitations of traditional approaches.
• To model image components, including occlusions and varying intensities, more effectively.
• To differentiate the impact of nonlinear versus linear modeling assumptions in sparse coding.

Main Methods:

• Developed a novel sparse coding model employing a spike-and-slab prior and a nonlinear max combination rule.
• Designed an exact Gibbs sampler for efficient inference in the intractable parameter optimization.
• Applied latent variable preselection for handling higher-dimensional data.

Main Results:

• The proposed model successfully extracts a sparse set of interpretable, edge-like components.
• The model's sparsity closely matches the ground-truth number of components in images.
• Demonstrated superior performance over linear models in learning edge-like components, especially with occlusion-rich data.

Conclusions:

• The novel spike-and-slab sparse coding model effectively captures image structures and occlusions.
• The nonlinear combination rule enables the representation of occluding image components.
• The model adaptively approximates and characterizes meaningful image generation processes, offering interpretable components.