What should the tokens be?

gaze direction/target?     gaze transitions?     image neighborhood codes?      - avoid explicit computation of gaze!