FROM BOTTOM-UP TO TOP-DOWN: CHARACTERIZATION OF TRAINING PROCESS IN GAZE MODELING

During training, artificial neural networks might not converge to a global minimum. Usually, using gradient descent, the training procedure cause the network to stroll in the high-dimensional weights? space. This stroll passes adjacently to local minima and locations in the geometry of loss landscape associated with low loss. Overall, the network moves from one low loss area to a lower loss area. In this work, we explored those low loss areas and minima, and tried to understand them. A U-Net was trained based on a gaze prediction task. A network was presented with images of different scenes, and the purpose of the network was to predict the expected human gaze distribution over those images. The driving task was selected since it involves relatively strong goal-oriented behaviors. It was shown that the training had two stages: (1) At the beginning, the network selected area was associated with saliency distributions (bottom-up behavior); (2) Later, the network selected area had the characteristics of goal-oriented distributions (top-down behavior) and it shifted away from the saliency distributions.