Table 1. Natural image model features and likelihood estimates.
BF | OS | DN | CP | Likelihood | References | |
(bits/pixel) | ||||||
RND/PCA/Whitening | x | 2.7 | [8], [43] | |||
ICA | x | x | 2.9 | [8], [43] | ||
-spherical | x | x | 3.05 | [8]–[10] | ||
-spherical | x | x | x | 3.17 | [62] | |
MEC with | x | x | x | x | 3.3 | [33] |
The natural image models we tested along with the neural response properties they mimic: “BF” is bandpass filtering, “OS” is orientation selectivity, “DN” is divisive normalization, and “CP” is complex cell pooling. We also show cited likelihood estimates. MEC is the mixture of elliptically contoured distributions model [33]. All models are described in detail in the “Models Tested” section. Higher likelihood indicates that a model captures more of the regularities present in natural images than a model with lower likelihood.