A Comparative Analysis of Visual Encoding Models Based on Classification and Segmentation Task-Driven CNNs

Algorithm 1

ROMP algorithm.

Input: observation matrix (specific features of a certain layer of CNN), observation vector (voxel responses), sparsity parameter ;

Output: weight vector ;

Process:

1. Initialization

Initialize the atomic support set , residual , and repeat the following steps times;

2. Atomic selection

Select the column index of the top maximum or all non-zero values (the number of non-zero coordinates is less than ) in , and form an atomic support set ;

3. Regularization

Find a subset in the set so that any two inner product and satisfy , and select the subset with the maximum energy among the subsets that satisfy the condition;

4. Update atomic support set and residual

. Update the residual: , and return to the second step.