Nonlinear Spike-And-Slab Sparse Coding for Interpretable Image Encoding
Fig 4
Parameter recovery on synthetic data.
Results of three differently parameterized sets of experiments, each with 10 experimental runs of 30 EM iterations on identical artificial ground-truth data generated according to the SSMCA model: A N = 2,000, D = 5 × 5. Three shown experimental settings are: B H′ = H = 10, C H′ = 5, and D H′ = 4, although the same results were obtained by the entire range of parameters H′ = [4, 10]. Importantly, the figure shows accurate recovery of ground-truth parameters which are plotted with dotted lines. B, C and D show in each column the parameter convergence of each of the three experiments, where the rows contain the following: data noise σ, sparsity H × π, prior standard dev. σpr, and the prior mean μpr. Finally, E shows the set of learned generative fields/components Wh corresponding to each experimental set B H′ = H = 10, C H′ = 5, and D H′ = 4.