Poster W47, Wed Dec 5, 7:30pm  12:00am

On sparsity and overcompleteness in image models
Pietro Berkes, Richard Turner, and Maneesh Sahani  Gatsby Computational Neuroscience Unit

The optimal linear model for natural images is very sparse but not overcomplete*
overcompleteness ratio marginal likelihood

sparseness,



sparseness,



*  This  result  applies  to  models  in  the  Studentt  family.  "Optimal"  means  that  the  marginal  likelihood  of  the  model  is  maximal.  "Very  sparse"  refers  to  the  optimal  degrees  of  freedom,  =2.09.  The  model  is  "not  overcomplete"  in  the      sense  that  the  optimal  number  of  dimensions  lies  between  1  and  1.3  times  the  number  of  input  dimensions.  Please  visit our poster for additional details.