Methods and systems are provided for training a neural network to track viewer engagement with a non-interactive display. Sensor data is obtained from one or more sensors, such as cameras, associated with a display device, and once a face is detected within the sensor data, a display sequence is initiated. The display sequence includes at least a first frame with a first visual feature and a second frame with a second visual feature. Using the sensor data obtained during presentation of the display sequence, viewer engagement with the sequence is tracked by determining eye movements and/or head movements. The detected eye movement and/or head movement is used to determine whether the person was actively engaged with the display sequence.