Generative Image Modeling

We use generative image models to study the probability distribution of natural images. For a recent review on the topic, see Gerhard et al. (2015). We put particular emphasis on an information theoretic quantitative evaluation.
model comparison
Comparison of several natural image models in terms of their ability to capture redundancies in natural images.
Generative image models have many applications ranging from density estimation to image reconstruction and texture synthesis. Computational neuroscientists are interested in image models because they are trying to understand to which extent the higher-order correlations present in natural scenes are exploited by neurons in the visual system. Redundancy reduction (Barlow, 1959), for example, is one of several principles the brain might use to learn image representations in an unsupervised manner. Generative image models allow us to derive algorithms which perform redundancy reduction (e.g., Sinz & Bethge, 2008).

One question of interest to us is how to best exploit hierarchical representations for image modeling. Hierarchical generative image models were outperformed by much simpler models for a long time (Theis et al., 2011), but recent advances are starting to change that (e.g., Theis & Bethge, 2015).

Selected References

L. Theis and M. Bethge
Generative Image Modeling Using Spatial LSTMs
Advances in Neural Information Processing Systems 28, 2015
#deep learning, #generative modeling, #natural image statistics, #lstm, #mcgsm
Code, URL, PDF, Supplemental, BibTex

H. E. Gerhard, L. Theis, and M. Bethge
Modeling Natural Image Statistics
Biologically-inspired Computer Vision—Fundamentals and Applications (to appear), Wiley VCH, 2015, ISBN 978-3527412648
#natural image statistics, #mcgsm, #ica, #psychophysics
URL, ISBN, PDF, BibTex

F. Sinz, J.-P. Lies, S. Gerwinn, and M. Bethge
Natter: A Python Natural Image Statistics Toolbox
Journal of Statistical Software, 61(5), 2014
#natural image statistics, #software, #python
Code, PDF, BibTex

L. Theis, S. Gerwinn, F. Sinz, and M. Bethge
In All Likelihood, Deep Belief Is Not Enough
Journal of Machine Learning Research, 12, 3071-3096, 2011
#natural image statistics, #deep belief networks, #boltzmann machines, #deep learning
Code, PDF, BibTex

R. Hosseini, F. Sinz, and M. Bethge
Lower bounds on the redundancy of natural images
Vision Research, 50(22), 2213-2222, 2010
#natural image statistics
PDF, BibTex

F. Sinz and M. Bethge
Lp-nested symmetric distributions
Journal of Machine Learning Research, 11, 3409-3451, 2010
#natural image statistics, #ica, #lp-spherically symmetric distributions, #nu-spherical symmetric distributions
PDF, BibTex

J. Eichhorn, F. Sinz, and M. Bethge
Natural Image Coding in V1: How Much Use Is Orientation Selectivity?
PLoS Computational Biology, 5(4), 2009
#natural image models, #natural image statistics, #normative models
Code, DOI, PDF, BibTex

F. Sinz and M. Bethge
The Conjoint Effect of Divisive Normalization and Orientation Selectivity on Redundancy Reduction
Advances in Neural Information Processing Systems 21, 2008
#contrast gain control, #normative models, #natural image statistics, #lp-spherically symmetric distributions
Code, PDF, BibTex

M. Bethge and R. Hosseini
Method and Device for Image Compression
Europäisches Patentamt, 2008
#mixture, #image compression
URL, BibTex
University of Tuebingen BCCN CIN MPI