Data-Efficient Representation Learning For Gaze Estimation