Natural images are very rare among the huge space of all possible images.

Uses a "pattern lexicon" of 5x5 images found to occur in a large set of natural images.

Matched filters: find the pattern with the closest match to the signal that was observed.

Frequency of Occurence: how many times the visual pattern appears in training material

5x5 selected because other sizes were deemed to have too many patterns or too expensive to compute.

Paper did not use overlapping convolution but just detected pattern for adjacent 5x5 tiles.

Tree is built by computer to reduce number of convolution operations needed to detect the pattern.

Notes

Paper was not as interesting as expected. Authors used personal judgement to determine which patterns were relevant and there is some manual tinkering involved.