Our lab mates were asked to draw reasonable copies of these training patterns using a mouse on a pixel editor interface. Columns B and C show examples of these mouse-drawn images. The images in columns B and C were correctly recognized by the system. Note that these images are not just a pure translation rotation or scaling of the training images. These examples show that the system demonstrates roubst invariance.

Column D shows another set of examples of mouse-drawn copies of training patterns in column A. These patterns were "incorrectly" recognized by the system. These patterns were recognized as a the patterns shown in  column E on the corresponding rows.