Figure 9. Example receptive fields learned with mixtures of natural scenes and noise (PoE model).

We created very sparse noise patterns, with only few pixels with substantial input, and mixed those in various proportions with natural scene input (or other distributions in G, H) for input to unsupervised learning. (A) Training the PoE model with 100% sparse noise resulted in highly-localized receptive fields. (B–F) Sparse noise continued to have a marked effect on the learned receptive fields even in the presence of natural scene input. With 50% natural input, receptive fields remained strongly localized (D), and even with 90% natural scenes, some pixel localization is still discernable (F). (G–H) This result was specific to sparse noise with a coefficient distribution near that of natural scenes. Training the PoE model with a mixture of natural scenes and either uniform white noise (G) or Gaussian (H) mixtures produced weaker perturbations of the receptive fields (cf panel D).